102 results for “TPU”

TextSeal2 is a distortion-free LLM text watermark that merges two or more keys into one virtual key and couples it to fresh noise through AutoMark's Gaussian copula. Under speculative decoding it verifies keyed drafts with an extra exponential race, so output matches the target model and acceptance rate is unchanged. On three LLMs at a 10^-3 false-positive rate, it detects 8 to 22 points more 100-token answers than TextSeal.

arxiv.orgpaperAdded 31 Oct 2026, published 29 Oct 2026🛡️ Security

TextSeal2 is a distortion-free watermark for LLM output that keeps its signal under speculative decoding. It merges two or more keys into one virtual key coupled to fresh noise through AutoMark's Gaussian copula, and verifies drafts with an extra exponential race. On three LLMs it detects 8 to 22 points more 100-token answers than TextSeal at a 10^-3 false-positive rate.

arxiv.orgpaperAdded 31 Oct 2026, published 29 Oct 2026🧠 AI

Understanding and Mitigating Role Spoofing in LLMs is an arXiv paper (October 2026) on role spoofing, a prompt injection where malicious tool output impersonates the user. It shows models track "who said what" using both content cues (semantics, style) and structure cues (control tokens, formatting). Attacks combining fake structure and content succeed often, adversarial training stops only content-based attacks, and linear probes trained to predict conversation roles detect spoofing across settings.

arxiv.orgpaperAdded 31 Oct 2026, published 29 Oct 2026🛡️ Security

This preprint proposes a unified post-training quantization framework for large language models. It derives a Hessian-guided approximation of the output change caused by joint weight and activation quantization, using a tractable Fisher approximation. Quantization ranges, linear transformations and weight rounding are optimized together under one output-sensitivity objective. The authors report gains over state-of-the-art transformation and rounding methods in channel-wise and block-wise settings.

arxiv.orgpaperAdded 31 Oct 2026, published 29 Oct 2026🧠 AI

SEAL is a defense against indirect prompt injection for local LLM agents. It learns a soft prompt and scoring readout offline, then uses a frozen task model to score and delete injected tokens from tool outputs, with queries built from trusted initial context. On AgentDojo, attack success falls from 34.6% to 0.1% and Joint rises from 51.4% to 77.5%, using about 127x fewer defense tokens than PISanitizer.

arxiv.orgpaperAdded 31 Oct 2026, published 29 Oct 2026🛡️ Security

High Precision Noise Cleaning for CKKS is an IACR ePrint preprint (2026/2675) by Zhihao Li, Cheng Hong and Tao Wei. It composes lower-degree Hermite polynomials to reach output noise bounds such as 2^-300 for noise flooding, and uses this to build an asynchronous threshold CKKS scheme with strong simulation security in the random oracle model. In OpenFHE, binary-message cleaning falls from 9.63 s to 1.54 s.

eprint.iacr.orgpaperAdded 31 Oct 2026, published 29 Oct 2026🛡️ Security

inkGround is a benchmark for stroke trajectory recovery from full-page, camera-captured handwriting, with 537 page images (about 18.7k words, 163k strokes) and pseudo-label references filtered by OCR agreement and human raters. Models get a page image and a target word, then output ordered stroke points. Seven VLMs were scored with DTW, Chamfer distance, readability and style similarity. Closed-source models, notably Gemini-3.1-Pro, did best, and the ranking held on IAM-OnDB.

arxiv.orgpaperAdded 30 Oct 2026, published 28 Oct 2026🧠 AI

The paper folds the per-key logit bias of Attention Matching (AM) KV cache compaction into the keys along one closed-form direction per head, fitted on context queries. The result is a plain cache that FlashAttention and vLLM paged caches can use, with output error bounded by how far each query's projection on that direction is from 1. Unmodified vLLM serves folded caches through a KV connector. On QuALITY and LongHealth the fold is within noise of AM in 13 of 14 settings. Under AM's published per-head budgets it is within noise in 9 of 10.

arxiv.orgpaperAdded 28 Oct 2026, published 27 Oct 2026🧠 AI

A stage-aware visual attribution framework for Vision-Language Navigation (VLN) that explains a robot's action choices. It identifies "instruction-stage insensitivity", where existing explainers give similar outputs across navigation stages. The method builds action-conditioned temporal instruction-visual attributions, aggregates relevant historical evidence by stage, and applies an instruction-guided spatial filter. It improves attribution quality across several VLN policies and benchmarks, helps diagnose failure cases, and was tested on a Unitree G1 robot.

arxiv.orgpaperAdded 27 Oct 2026, published 25 Oct 2026🧠 AI

FORK is a reward for rubric-based reinforcement learning that jointly optimizes recall and a new precision objective. Precision requires each unit of a response to be necessary (rubric-driven necessity credits) and factually correct (a grader). Rubric-coverage rewards otherwise produce verbose, less accurate responses. FORK keeps similar coverage with shorter outputs, raising win rate against the base policy from 10.6% to 68.0% on RubricHub-Medical and from 22.1% to 74.1% on HealthBench.

arxiv.orgpaperAdded 27 Oct 2026, published 24 Oct 2026🧠 AI

GordieHowe is a language, certifying compiler and zero-knowledge proof system for building anonymous credentials on standardized signatures (ECDSA, RSA, ML-DSA). It proves an ECDSA-signed mobile driving license presentation in 64.8 ms with a 276 KiB proof, 10.1x faster than Google's Longfellow, and an ML-DSA-44 signature in 87.3 ms with at least 116 bits of quantum security. Compiler output and security theorems are checked in Lean. It introduces series-parallel IOPs and extends the BCS transformation in the QROM.

eprint.iacr.orgpaperAdded 24 Oct 2026, published 23 Oct 2026🛡️ Security

Pairwise Independence of Ten-Round AES (Zhongfeng Niu, Kai Hu, ePrint 2026/2608) proves that ten rounds of AES with independent uniform round keys and input/output whitening are within 3017/2^141 (below 2^-129) statistical distance of pairwise independent. The proof bounds the averaged difference-distribution matrix over five superbox layers, using Gram moments and MDS code constraints. Final-round MixColumns omission is retained.

eprint.iacr.orgpaperAdded 24 Oct 2026, published 23 Oct 2026🛡️ Security

The Geometry of Metaphor is an October 2026 arXiv preprint that uses mechanistic interpretability to find causal subspaces for metaphor in LLMs of different architectures and sizes. The authors steer output on classification tasks across parts of speech. They also steer generation tasks, showing the metaphoricity of model output can be controlled. The abstract gives no model names or metrics.

arxiv.orgpaperAdded 24 Oct 2026, published 22 Oct 2026🧠 AI

RIFT is a family of selection-based KV cache compression strategies for speculative-decoding draft models. It targets "decode staleness", where prefill-focused selection loses relevance as output grows, by balancing prefill information with recent generated tokens, refreshing the selection adaptively, and training drafters with sparse attention masks. Qwen3.5-4B drafting for Qwen3.5-122B keeps over 98% of full-cache acceptance length at 32K input, 4K output and an 8K KV budget.

arxiv.orgpaperAdded 23 Oct 2026, published 22 Oct 2026🧠 AI

This IACR ePrint paper (2026/2597) systematically compares ring-oscillator-based additive noise generators, a hiding countermeasure against side-channel attacks. Experiments show many existing designs ignore output amplification and the risk of the noise being filtered out, so they lower signal-to-noise ratio too little. The authors give a guideline on inverter chain length, output amplification and frequency randomization, and build a tailored generator for an AES implementation that outperforms prior designs.

eprint.iacr.orgpaperAdded 24 Oct 2026, published 22 Oct 2026🛡️ Security

Monitorability Is Trainable proposes an unsupervised post-training objective that maximizes mutual information between a model's chain of thought and its output. On three model organisms with degraded CoT (from compression, monitor optimization, and synthetic-document generalization), it recovers most or all lost monitorability. On undegraded policies it improves CoT monitor and activation probe prediction of unsafe behavior and raises consequential reasoning.

arxiv.orgpaperAdded 23 Oct 2026, published 21 Oct 2026🧠 AI

Cache Engineering for Agents proposes letting LLM agents control their active KV cache directly, using a drop operation to exclude cache from attention and repos to adjust retained keys' position encodings, without editing the trajectory or re-prefilling. The authors build the AKV serving engine to handle history-dependent cache states. A tool-response rolling-drop policy raises average benchmark scores by 1.7 points and cuts cost by 27.7% against threshold-triggered summarization. AKV gives up to 1.2x output throughput and 1.5x faster P95 latency.

arxiv.orgpaperAdded 22 Oct 2026, published 20 Oct 2026🧠 AI

This paper shows that reward misspecification alone degrades chain-of-thought (CoT) monitorability in language models. Using a testbed of paired true and misspecified rewards across three models and three tasks, CoTs keep reasoning about task intent while outputs hack the reward. The proposed reward description training injects an LM-inferred reward description into the input and raises monitorability by an average of 24% or 42% absolute, depending on the metric.

arxiv.orgpaperAdded 22 Oct 2026, published 20 Oct 2026🧠 AI

This paper studies how keeping long tool output out of the context window, and reading it through a section index instead, costs LLM agents accuracy. Using the PHREEQC geochemistry simulator on 158 execution-verified multiple-choice items and 11,007 logged calls, GPT-5.1 and GPT-5.2 lose 11.3 and 7.9 points, while Claude Opus 4.6 and Sonnet 4.6 show no detectable cost. The authors name the failure evidence substitution: agents skip the report and misround from their own output files.

arxiv.orgpaperAdded 21 Oct 2026, published 19 Oct 2026🧠 AI

EpiLens is a lightweight, open-source retrieval-augmented generation package that applies a fixed retrieval template and output schema to every paper in a corpus and links outputs to source passages. On a corpus of 2,297 epidemiology papers, its classifier matches or exceeds supervised and zero-shot baselines on four tasks without task-specific training data, scored against 214 manual labels. It also compares model-based study shares across parameters and pathogens.

arxiv.orgpaperAdded 21 Oct 2026, published 19 Oct 2026🧠 AI

SpanGuard is an indirect prompt injection (IPI) guardrail that splits detection into two steps. Lightweight unsupervised segmentation first locates candidate instruction blocks in tool outputs, then a small LLM judges whether each block is authorized by the user's request. The authors report a 0.32% attack success rate with negligible benign utility loss, and the guardrail still works with substantially smaller models.

arxiv.orgpaperAdded 21 Oct 2026, published 19 Oct 2026🛡️ Security

The paper presents a fault attack on deterministic CSIDH and OSIDH-LD. Forcing the lane-selection mask to zero (for example by skipping one instruction) makes output changes reveal whether a secret exponent reaches a threshold. Binary search then recovers the full key in at most 297 queries for OSIDH-LD and 222 for deterministic CSIDH with CSIDH-512-sized keys. Full OSIDH-LD keys were recovered in software fault simulation, and a ChipWhisperer-Lite skipped the target instruction with probability at least 0.94. The authors argue fault resistance requires giving up determinism and propose randomization countermeasures.

eprint.iacr.orgpaperAdded 20 Oct 2026, published 19 Oct 2026🛡️ Security

"How employment is changing at Anthropic" is an October 2026 technical report by Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory. It uses internal employment, agent telemetry, output, survey, calendar and hiring data plus 17 interviews. Junior technical hiring fell from 27% to 9%, agent hours reached six times employee hours by August 2026, and adopters' output rose. Employees increasingly do meta-work managing agents.

anthropic.compaperAdded 19 Oct 2026, published 18 Oct 2026🧠 AI

Doc2Harness is a two-stage benchmark for harness engineering. In Stage 1 a build model, working inside a fixed builder harness, gets an interface-only contract skeleton plus a development document consolidating behavior from twelve frontier harnesses, and builds a full candidate harness in an isolated container. In Stage 2 the frozen harness runs in separate per-test-case containers, driven by a scripted mock model, and is scored on four channels: requests sent to the model, files on disk, exit code and command output. Across 19 frontier models with three builds each on 328 test cases, GPT-6 Astra leads at 53.05% against 27.44% for the weakest, and failures cluster around how cross-cutting functionality is split across modules.

huggingface.copaperAdded 16 Oct 2026🧠 AI

DK-VLM adds Dual-Kernel Attention to a vision-language model. A semantic kernel keeps the pretrained attention and positional encoding. A separate geometric kernel uses estimated 3D points and camera poses to modulate attention through point displacements and relative view transformations, and a learned gate fuses the two outputs. Geometry therefore enters attention as position, not as fused 3D reconstruction features. The paper also adds Reference-Frame Composition (RC) supervision, which gives stepwise turn and heading targets derived from verified single-step annotations, to train multi-step spatial reasoning as the observer moves. DK-VLM-8B reports the highest average score across six spatial benchmarks among the compared methods, with 74.33 on VSI-Bench and 58.28 on ReVSI-32, and it preserves general video understanding.

arxiv.orgpaperAdded 17 Oct 2026, published 15 Oct 2026🧠 AI

DropKV is a KV-cache eviction framework for long-context LLMs. It decouples the joint eviction decision into independent per-token scores based on residual-output perturbation, which avoids the NP-hard combinatorial problem. It has a provable constant-factor approximation guarantee under a cone condition, which the authors verify empirically on three long-context LLMs. The fused Triton kernels never materialize the full attention matrix. They give a 19.8x scoring-kernel speedup and up to 9.9% end-to-end prefill speedup. On RULER and LongBench it beats state-of-the-art eviction baselines at equal cache budgets.

arxiv.orgpaperAdded 17 Oct 2026, published 15 Oct 2026🧠 AI

Evaluates a native probabilistic decision model (Jev), Llama Prompt Guard 2 (22M and 86M) and an LLM judge on 47,611 held-out examples covering direct/indirect prompt injection, hard-benign inputs, obfuscation and distribution shift. The LLM judge has the best AUROC (0.897), followed by Jev (0.871), Prompt Guard 2 86M (0.757) and 22M (0.714). Jev's native probabilities do not consistently beat post-hoc calibrated classifier scores, and temperature scaling improves calibration but not discrimination. At a 1% false-negative-rate target, every system needs over 91.8% escalation, so probabilistic outputs alone do not give reliable security decisions under distribution shift.

arxiv.orgpaperAdded 17 Oct 2026, published 14 Oct 2026🛡️ Security

ChunkAct brings action chunking to agentic robotics, where an LLM agent composes perception, motion and VLA tools guided by memory. Online, one call launches a chain of primitives that runs autonomously and returns control only on failure. Offline, memory is compiled into these chunks so execution needs no memory at test time and the planner is used only for real decisions. On LIBERO-Pro it improves task success by 5.0 points over the prior agentic-robotics state of the art, with 2.9x fewer LLM calls, 4.4x fewer input tokens and 2.3x fewer output tokens.

arxiv.orgpaperAdded 14 Oct 2026, published 13 Oct 2026🧠 AI

RepProbe is a benchmark of 42K real and synthetic video clips with exactly 1–20 complete action repetitions, controlling for clip position, temporal coverage and duration. Video language models' counting accuracy collapses at high counts, and this is not explained by frame sampling alone. Linear probes on final-layer representations show count information stays linearly decodable even when the model's answer is wrong, so the failure arises at the output stage, where answers cluster on a few preferred numbers.

arxiv.orgpaperAdded 14 Oct 2026, published 13 Oct 2026🧠 AI

EmoGenie is a benchmark for emotions expressed and evoked by text-to-image outputs. It has 6,400 images from eight T2I models using 800 emotion-targeted prompts. It includes EmoGenie-QA (90,000 closed-set and 1,057 open-ended questions) and EmoGenie-Assess (13,175 mean opinion scores for valence, arousal, dominance and aesthetic quality). The authors also propose EmoProbe, which pairs emotion QA with response-aligned, dimension-specific regression heads. They report that it beats baselines on emotion QA and correlates better with human ratings. The work shows how reliably MLLMs interpret affect in generated images.

arxiv.orgpaperAdded 14 Oct 2026, published 13 Oct 2026🧠 AI

Marcel Keller's ePrint compares the two fully secure MPC sorting approaches, Quicksort (Araki et al., CCS'21) and radix sort (Asharov et al., CCS'22). The comparison covers a range of parameters, for both honest- and dishonest-majority settings. The paper then extends the sparse matrix multiplication of Damie et al. (CODASPY'26), which avoids oblivious RAM, so the output comes out in the same format as the inputs. That makes repeated multiplication of graph adjacency matrices possible, which the paper uses to find cycles and scatter-gather patterns associated with money laundering. The result is privacy-preserving anti-money-laundering detection.

eprint.iacr.orgpaperAdded 14 Oct 2026, published 13 Oct 2026🛡️ Security

StableJudge is a defense framework against score-hijacking prompt injection in LLM-as-a-Judge pipelines (e.g. paper review, recommendation). A compact 0.13B "Sentry" model localizes injected spans and sanitizes them before the judge scores the content, trained on adversarially evolving attacks, and deterministic fine-grained scores are derived from the judge's output distribution. It improves score recovery under attack on held-out scientific abstracts and out-of-distribution attack variants while preserving general performance, decoupling safety from judge model scale. Code and data are promised but not yet released.

arxiv.orgpaperAdded 15 Oct 2026, published 12 Oct 2026🛡️ Security

Presents an EM side-channel attack on LLaMA-3.2-1B-Instruct running on a Raspberry Pi 5, showing passive physical measurements leak generated tokens and user prompts. Three attacks: a profiling attack (97.2% accuracy over 50 medical/privacy keywords), a hidden-state reconstruction attack predicting tokens from the 128k vocabulary (94.3% on profiled classes; 67% of unseen words ranked in top 100), and a hidden-state inversion attack reconstructing prompts semantically. Shows that local edge inference does not guarantee confidentiality of tokens and inputs.

arxiv.orgpaperAdded 16 Oct 2026, published 9 Oct 2026🛡️ Security

Hide and Seek (HnS) is a concept-level attribution method for black-box vision models. It needs no model weights and no concept-labeled datasets. Given a concept and an image set, it uses a vision-language model to retrieve concept-aligned images and localize the concept, masks the localized region, and takes attribution from the change in the black-box model's output. It agreed better with ground-truth attributions than existing methods on CelebA and COCO-Stuff, and it works on tasks existing concept attribution methods cannot handle, such as image captioning and satellite image segmentation.

arxiv.orgpaperAdded 12 Oct 2026, published 9 Oct 2026🧠 AI

Introduces EICS (Evaluation-Invariant Concept Stability), a metric for checking whether representations consistently encode a concept under controlled one-concept interventions. Also introduces CASP-VL, a benchmark of causally controlled spatial perturbations derived from Visual Genome scene graphs. Across several VLM architectures, spatial relations are represented much less faithfully than lexical information, even for models well above chance on compositional benchmarks. The relation is identifiable at the predicate's own tokens but is lost when the sentence is aggregated into a global representation (EOS token or mean pooling), a failure that output-based evaluation does not reveal.

arxiv.orgpaperAdded 12 Oct 2026, published 9 Oct 2026🧠 AI

Cryptanalysis of the Alternating Sponge (ASP) and Duet, a permutation-based hash candidate in China's ICCS Next Generation Commercial Cryptographic Algorithm program (NGCC). The authors give an O(2^r)-query distinguisher against ASP that contradicts its claimed indifferentiability bound in some parameter regimes. They also give a distinguisher and state-recovery attack on Duet's multi-block output, using O(2^(r+c-h)) permutation queries. Collision and preimage resistance are unaffected, but Duet used as a XOF or KDF is affected, and state recovery enables length extension.

eprint.iacr.orgpaperAdded 12 Oct 2026, published 9 Oct 2026🛡️ Security

ReplaySSM represents state-space model state as a periodically updated checkpoint plus a buffer of recent SSM inputs, separating output computation from checkpoint updates. This avoids state writes on most decoding steps (roughly halving dominant state memory traffic), makes speculative-decoding rollback a buffer-pointer operation, and lets a low-precision checkpoint serve outputs while a higher-precision one drives recurrent updates. Implemented in vLLM for Mamba-2 and Gated DeltaNet, it gives 1.20–1.47x decoding speedup at batch size 256 across four hybrid models and about 1.8x the throughput of vLLM speculative decoding on two 120B-class models.

arxiv.orgpaperAdded 12 Oct 2026, published 9 Oct 2026🧠 AI

Adaptive DHiSparse extends HiSparse (host-memory KV storage that fetches only selected entries) to draft-model speculative decoding methods DSpark and DFlash2. It protects the required working set, keeps extra history while GPU memory is spare, and recovers the rest by exact fused retrieval. Queries share missing-row retrieval, draft KV uses a window-bounded ring, and verification staging is reused across layer groups. At 100K context and concurrency 128, GLM-5.3 with DFlash2 reaches 793 output tokens/s per decode GPU (1.45x HiSparse). With the same latency bounds it sustains up to 4x the concurrency. On DeepSeek-V4.1-Flash at 500K context, it delivers 1.85x the output of resident serving with DSpark.

arxiv.orgpaperAdded 12 Oct 2026, published 9 Oct 2026🧠 AI

Proposes the Answer-Evidence Support Protocol (AESP), which checks whether the grounding evidence a model outputs actually supports its own answer in grounded remote sensing visual question answering. It applies task-specific rules to existence, counting and comparison questions. The authors build the G-RSVQA benchmark of 15,027 samples with questions, reference answers and object-level annotations. Tests on seven grounded vision-language models show evidence-support failures on all three question types, with up to 50.82% of correct answers unsupported by their grounding evidence.

arxiv.orgpaperAdded 12 Oct 2026, published 9 Oct 2026🧠 AI

Proposes "study contracts" that bind an AI-generated study's declared experimental choices, run obligations and claim scope to recorded execution evidence, defining contract-relative verification as distinct from scientific truth. In a diagnostic with eight self-authored clean/mutated pairs, a deterministic fault-specific checker caught all eight mutations, while eighteen LLM judge aliases given metadata-filtered packages flagged 104 of 144 mutated cases (32 abstentions, 8 terminal failures). The authors stress the setting is deliberately information-asymmetric, not a verifier-quality comparison or reward-hacking result, and call for full-information comparisons, legitimate-adaptation controls and closed-loop agent evaluations.

arxiv.orgpaperAdded 9 Oct 2026, published 8 Oct 2026🧠 AI

Proposes making the verifier the evolving object in self-improving agent loops. The verifier is an inspectable expression over small, mostly deterministic drawback detectors, synthesized from clustered failures, gated at birth, and selected for agreement with a ten-item anchored reference set plus consensus over unlabeled outputs, never for the agent's score. On MBPP+ it gains +0.21 held-out agreement over the hand-authored seed and ends ahead of the bare LLM judge it contains. Removing the anchor guards collapses the verifier into an always-pass grader that still trains skills as well, so downstream task score cannot certify a self-evolved verifier. The paper's Double Ratchet pairs the verifier with a lifecycle-managed skill loop and retains 88-110% of the lift from ground truth or a rubric on code generation, enterprise text-to-SQL and reference-free report generation.

arxiv.orgpaperAdded 9 Oct 2026, published 8 Oct 2026🧠 AI

SLDR is a post-fine-tuning defense against malicious fine-tuning-as-a-service attacks, where an attacker fine-tunes an aligned LLM on a legitimate-looking task to covertly erode its refusal behavior. The method trains a LoRA "recovery" adapter only on the layers found to have maximal and minimal signed safety sensitivity, and uses representation-based dynamic routing at inference time to activate that adapter only for malicious queries, leaving benign task performance untouched. On Llama3.1 fine-tuned for SST2, it cuts the average harmful-output score from 11.54 to 0.08 while keeping downstream accuracy, and stays near-zero even when up to 90% of the fine-tuning data is poisoned.

arxiv.orgpaperAdded 9 Oct 2026, published 7 Oct 2026🛡️ Security

The paper introduces DIGHT, a co-adaptive framework that couples a text-conditioned digital-human interaction generator with a physics-based humanoid tracking policy to produce physically plausible two-humanoid interactions. Rather than collapsing executability and interaction-fidelity signals into one scalar reward, it applies physics-decoupled diffusion Direct Preference Optimization (DPO) to align the generator using criterion-specific preferences derived from tracking error, friction, floating, and simulator force feedback (contact occurrence, location, duration, magnitude), then uses the aligned generator's outputs to fine-tune the tracker. Experiments show improved physical plausibility and more reliable, faithful simulated humanoid interactions compared to prior tracking-based approaches.

arxiv.orgpaperAdded 8 Oct 2026, published 7 Oct 2026🧠 AI

This paper argues that training "model organisms" (models deliberately given alignment-relevant behaviors like backdoors, sycophancy, or spurious correlations) to a single behavior-installation objective produces unrealistic test subjects for interpretability research. The authors propose a three-objective validation framework (behavior installation, general-capability preservation, output naturalness) and show chat quality and chain-of-thought naturalness degrade substantially across common training recipes, with these degradations predicting how well interpretability methods such as logit lens recover the installed behavior. They introduce a model-merging-based multi-objective training approach and apply it to a new model-organism suite for demographic bias in clinical reasoning, finding DPO training preserves the base model better than supervised finetuning.

arxiv.orgpaperAdded 8 Oct 2026, published 7 Oct 2026🧠 AI

This paper introduces Alhena (synchronous) and Wasat (asynchronous), two helper-aided multi-party computation protocols that achieve "identifiable fairness": if an adversary aborts the protocol, no party gets the output and honest parties can identify at least one common cheater, closing a gap in prior fair-but-unaccountable helper-aided MPC schemes. The authors benchmark against state-of-the-art fair protocols Asterisk and Castor (both from Kamarkar et al., IEEE S&P 2024/2026), showing 4–12.5x and 2.5–5x online-phase speedups respectively, with online performance unaffected by active misbehavior.

eprint.iacr.orgpaperAdded 8 Oct 2026, published 7 Oct 2026🛡️ Security

World Models' Last Exam in Physics introduces a measurement-based benchmark of 40 controlled tasks across mechanics, optics, fluids, thermal/phase-change phenomena, electromagnetism, and surface tension, designed to test whether video world models generate physically consistent outputs. Each task pairs an initial image and prompt with predefined physical criteria, and a custom evaluator combines observability screening with task-specific quantitative measurements rather than relying on reference videos or model-based judgments. Testing eight video generation models across 1,280 generated videos found persistent physical inconsistencies, with the best model scoring only 57.76/100, and the evaluator showed higher agreement with human judgments than a vision-language model baseline.

arxiv.orgpaperAdded 7 Oct 2026, published 6 Oct 2026🧠 AI

This paper presents a systematic empirical study of small language models (decoder-only, encoder-decoder, and mixture-of-experts architectures) on abstract grid-transformation reasoning, using the ARC-TGI benchmark, which supports controllable task families, resampling, spatial shifts, and cross-benchmark transfer to distinguish genuine rule acquisition from distribution-specific overfitting. Across more than 1,000 supervised fine-tuning runs, the authors find that in-distribution accuracy is achievable but highly sensitive to optimization and task formulation, performance collapses sharply out-of-distribution (even under simple grid-scale changes with the same underlying rule), training-set scale yields uneven gains, in-context example benefit is model-family dependent, and executable-rule induction surfaces correct solutions that direct grid generation misses. Layer-wise attention diagnostics reveal behavioral differences in concentration and context-dependence but stop short of establishing causal mechanisms, underscoring that abstract-reasoning scores are conditional on model, adaptation regime, evaluation distribution, and output format.

arxiv.orgpaperAdded 7 Oct 2026, published 6 Oct 2026🧠 AI

This paper introduces Semantic Behavioral Watermarking (SBW), a method for embedding an owner identifier into an LLM agent's high-level tool/action choices rather than output tokens, so provenance survives tool renaming and observation paraphrasing. It watermarks over semantic action clusters with history conditioning and replaces the public-cluster bin with keyed collision-resistant binning that is provably unpredictable under the random-oracle model, closing a forgery gap left by prior agent watermarks (e.g., AgentMark, whose bit-recovery drops to 16.8% under paraphrasing). Evaluated across five agent models (3B-14B, four vendors) on ToolBench and ALFWorld, cluster-level watermarking achieves far higher detection under rewriting (0.49-0.66 vs 0.05-0.17 on ToolBench; 0.92-0.97 vs 0.00-0.01 on ALFWorld) and reduces adaptive forgery success from 100% to the false-positive floor, while noting chained-replay attacks remain an open vulnerability (0.76-0.98 detection failure).

arxiv.orgpaperAdded 7 Oct 2026, published 6 Oct 2026🧠 AI

OVAL (Output-Aware Local page bases) is a training-free method for encoding KV cache pages in page-sparse attention, designed so that retrieval directly accounts for how approximation errors affect the value-weighted attention output rather than just estimating attention scores. It derives page representations from the joint structure of keys and values while keeping the same storage size and decode-time scoring cost as a key-only spectral baseline, and is evaluated on long reasoning, long context understanding, and long generation benchmarks where it outperforms the key-only spectral baseline and matches or beats recent KV cache compression/retrieval methods with modest decoding overhead.

arxiv.orgpaperAdded 6 Oct 2026, published 5 Oct 2026🧠 AI

This arXiv paper introduces RAISED (Robust Attack Invariance through Self-Distillation), a training framework that defends tool-using LLM agents against indirect prompt injection without the utility degradation seen in prior training-based defenses. The authors show existing defenses cause distributional drift that makes models wrongly refuse legitimate tool-output-guided steps on benign tasks, then fix this by having a model generate its own tool-use scenarios and self-distill to match clean-context behavior on both clean and injected trajectory variants. Results show reduced attack success rate on injected tool responses while preserving performance on agentic and general-purpose benchmarks.

arxiv.orgpaperAdded 6 Oct 2026, published 5 Oct 2026🛡️ Security

An IACR ePrint paper presenting a GPU-assisted implementation of FAEST v3 signing, a post-quantum signature scheme, focusing on two problems: preserving the correct minimum accepting counter when parallelizing "challenge grinding" across GPU threads, and building a two-slot CPU-GPU pipeline to overlap CPU work with GPU batches for multi-signature throughput. Across the six non-EM FAEST v3 parameter sets, the authors find the CPU-GPU path beats optimized CPU-only signing for single-sign latency in 4 of 6 cases (up to 1.6x speedup for FAEST-192s) and wins throughput in 9 of 18 parameter/batch-size combinations, with GPU correctness validated to match the reference signature output.

eprint.iacr.orgpaperAdded 8 Oct 2026, published 5 Oct 2026🛡️ Security

This IACR eprint paper studies output policies for CKKS homomorphic-encryption-based anomaly detection, evaluating whether limiting released scores and decisions can hide a private detection threshold without breaking existing decisions. Across 45 plaintext and 45 CKKS experiments on three intrusion-detection datasets and autoencoder scoring models, the authors show thresholds remain estimable, quantization preserves about 99.86 percent of decisions, and a surrogate model trained on only 128 decision-only responses achieves 0.817 mean balanced accuracy at predicting decisions on new inputs, a decision-prediction model-extraction-style attack whose success depends heavily on data-splitting strategy. It argues numerical threshold leakage and decision-prediction leakage are distinct protection goals that output policies for encrypted intrusion detection systems must address separately.

eprint.iacr.orgpaperAdded 8 Oct 2026, published 5 Oct 2026🛡️ Security

CyTReX is a proposed explainable-AI framework for Distributed Energy Resource (DER) network security that converts edge anomaly-detector alerts into ranked, evidence-traceable threat hypotheses for SOC analysts. It builds a structured "evidence packet" combining edge anomaly detection outputs, cloud-layer attack interpretation, SHAP feature attributions, surrogate decision rules, and Model Context Protocol-enabled cyber threat intelligence enrichment, constraining LLM reasoning so every ranked hypothesis and attack-tree branch is traceable to explicit evidence rather than free-form inference. The authors evaluate five pipeline configurations and show that adding reasoning components improves hypothesis specificity and evidence traceability for SOC triage/investigation.

arxiv.orgpaperAdded 6 Oct 2026, published 3 Oct 2026🛡️ Security

LESSER is a drop-in wrapper for gradient-based post-training data selection that approximates full-parameter gradient alignment using only output-layer gradients, which can be computed with a forward pass instead of a full backward pass. It cuts feature-extraction compute by 9.7x on SFT benchmarks and 3.0x on RL benchmarks while matching the downstream performance of full-gradient selection methods, even though individual sample rankings differ, selected batches remain gradient-aligned.

arxiv.orgpaperAdded 5 Oct 2026, published 2 Oct 2026🧠 AI

This paper shows that training policies against monitors (an in-domain activation probe and two penalties that score how early a model commits to its final answer) can drive the monitor's readout to a numerical floor without actually controlling the underlying reward-hacking behavior. In a code-generation RL environment where the exploit is available from the start of the reasoning trace, runs with an identical zero-median "commitment" score range from mostly honest to near-pure reward hacking depending only on random seed, and probe-trained runs reach the hacking regime despite floor-level readouts, because models learn to insert generic planning/filler text that delays the exploit past the point the monitor reads without removing it from the output. The authors argue offline monitor validation plus a low training-time readout is not sufficient evidence of behavioral control and that an out-of-band behavioral check is required.

arxiv.orgpaperAdded 5 Oct 2026, published 2 Oct 2026🧠 AI

This arXiv paper replays ground-truth tool calls from the AgentDojo and tau-bench agent benchmarks (without an LLM) to generate benign-by-construction tool outputs, then labels injected outputs via differential replay to evaluate fifteen prompt-injection detectors, including Meta's Prompt Guard 2, plus two task-aware LLM judges, also testing them on BIPIA. It finds detection rankings transfer poorly between benchmarks (the best detector on BIPIA catches only 2% of AgentDojo injections at 1% false-positive rate; a detector catching 72% of AgentDojo injections catches only 15% on tau-bench), while false-positive rates do transfer between the two agent benchmarks, and traces this to mismatches between training data format (e.g., training on short prompt strings vs. tool-output-embedded injections) and deployment context.

arxiv.orgpaperAdded 5 Oct 2026, published 2 Oct 2026🛡️ Security

Persona Guardrail is a production-grade runtime defense framework for customer-facing LLM agents that enforces explicit functional boundaries via synchronous input/output validation against semantic allowlist and blocklist specifications. The authors also introduce PAGE (Persona-Aware Guardrail Evaluation), a benchmark covering benign, adversarial, and out-of-domain user and agent turns; deployed in production, it raises overall accuracy from 85.7% to 95.9% versus a generic LLM-based guardrail, boosts out-of-domain detection from 57.3% to 93.5%, and cuts the false-approved rate from 25.0% to 4.7%, all within latency budget.

arxiv.orgpaperAdded 5 Oct 2026, published 2 Oct 2026🛡️ Security

An arXiv preprint that empirically tests speculative decoding on consumer hardware (Intel Arc B570, NVIDIA RTX, Apple M4-class systems), distinguishing between 'execution-path qualification' (whether a speculative predictor's output matches the target model's exact generation trajectory) and the realized latency/cost of running speculation. The authors find that a target-only, draft-free predictor matches greedy output on most repeat-stable prompts but does not guarantee qualification or identify a unique cause of divergence, and that newly qualified RTX serialization and other configurations (358H, M4 Guard) are often slower than target-only decoding once real costs are counted. The paper's conclusion is that these results are diagnostic engineering findings specific to the tested hardware/software paths, not evidence of a portable decoding speedup.

arxiv.orgpaperAdded 5 Oct 2026, published 2 Oct 2026🧠 AI

An arXiv thesis-style paper proposing R_stab(f), a Jensen-Shannon-divergence-based robustness metric for LLM outputs under small input perturbations, along with a proven bound V(h) <= 1 - R_class(h) for localized attacks. It introduces ASA, an adaptive evolutionary black-box attack against LLM-as-a-Judge systems reaching up to 73.8% attack success rate (62.6% cross-model transfer), evaluates trojan/backdoor detection on the Trojan Detection Challenge 2023 (Pythia-1.4B) and SaTML CTF 2024, shows that committees of 5-7 heterogeneous models cut Gemma-3-4B's ASR to 19.3%, and presents AttestMCP, an HMAC-based tool-call attestation scheme plus a "Commit Boundary" isolation pattern for MCP agents that cuts average ASR on the 847-scenario MCPBench from 53.7% to 12.4%. Methods ship as the JudgeGuard, TrojanArmor, and MCPSec software suites.

arxiv.orgpaperAdded 5 Oct 2026, published 1 Oct 2026🛡️ Security

This paper studies whether AI oversight mechanisms based on interactive proofs and debate (used for verifying AI outputs derived from confidential data, e.g. a fitness-for-duty assessment from medical records) can be made zero-knowledge, so a verifier learns nothing beyond correctness. The authors prove a general impossibility result in the random oracle model — no zero-knowledge proofs exist for general oracle-aided computation, and this extends to debate, a canonical scalable-oversight model — but show that if the oracle cryptographically signs its answers, every oracle-aided computation can be verified in zero knowledge with an efficient prover and verifier using only collision-resistant hash functions. The signed-oracle construction also offers a new approach to scalable oversight that avoids relying on an honest debate opponent or on the computation's robustness.

arxiv.orgpaperAdded 2 Oct 2026, published 1 Oct 2026🧠 AI

This arXiv paper presents a reproducible benchmark that unifies evaluation of six black-box LLM extraction attacks, ten defenses, and two adaptive attacks (paraphrasing and back-translation of protected responses) under controlled query budgets and model configurations. It measures surrogate capability, fidelity to the victim model, output quality (Rep-4), and provenance-detector scores, giving a consistent basis for comparing how extraction methods and defenses interact under text-only API access. Code and artifacts are linked from the paper.

arxiv.orgpaperAdded 3 Oct 2026, published 30 Sep 2026🛡️ Security

ViTeX-Bench is a benchmark suite for high-fidelity video scene text editing, pairing a dataset of 387 real-world 720p videos (text-region masks plus editing instructions, 230 for training and 157 frozen for evaluation) with a three-axis, 13-metric evaluation protocol covering text correctness, visual/temporal quality, and edit locality. The authors benchmark eight baselines across four editing families and release ViTeX-Edit-14B, an open-source reference editor using motion-aligned glyph-video conditioning that achieves the best CharAcc (0.688) among video-native editors and the lowest text-crop Warp among raw outputs.

arxiv.orgpaperAdded 1 Oct 2026, published 30 Sep 2026🧠 AI

This paper studies whether weight tying between input and output embeddings, common in decoder-only LLMs for parameter efficiency, remains beneficial when models are fine-tuned under Differentially Private Stochastic Gradient Descent (DP-SGD). Using GPT2 and DistilGPT2 fine-tuned on SST-2, QNLI, and QQP, the authors find untied embeddings consistently outperform tied ones by up to 4.74 accuracy points, and that untying also enables memory-efficient ghost clipping (since tied embeddings create shared-parameter interactions that break standard ghost norm computation), yielding over 60% lower memory usage. The results suggest standard LLM architectural choices should be revisited for privacy-preserving training settings.

arxiv.orgpaperAdded 1 Oct 2026, published 30 Sep 2026🧠 AI

This IACR ePrint paper introduces zkPoPE (Zero-Knowledge Proof of Protocol Execution), the first transport-agnostic security model for zkTLS, formalized as a UC ideal functionality that separates "proof of an actual protocol execution occurred" from a plain NIZK over input-output pairs. The authors formally analyze the Reclaim Protocol, showing it UC-realizes zkPoPE under a deployment-aligned trust model, and introduce Distributed-AEAD, a new three-party authenticated-encryption primitive enabling Reclaim to attest a 1200kB TLS response on a lower-mid-range phone in 7.28s — two orders of magnitude larger than previously benchmarked zkTLS schemes. It matters because it gives zkTLS constructions (used for privacy-preserving web attestations) a common, construction-independent security definition to be proven against, rather than each scheme inventing its own.

iacr.orgpaperAdded 29 Sep 2026🛡️ Security

This paper shows that adding an adversarial loss as a post-training correction to pretrained pixel diffusion/flow-matching models (applied at non-high-noise timesteps, architecture and sampler unchanged) restores missing high-frequency, fine-scale natural-image statistics that these models systematically underproduce. Through frequency-band/power-law analysis and nearest-neighbor, recall, and no-GAN matched-SFT controls, the authors rule out memorization or mode-dropping as explanations, and show the effect jointly improves distribution fidelity, coverage, prompt alignment and perceptual quality — but does not transfer comparably to latent diffusion configurations, pointing to direct pixel-output access as the key enabling factor.

arxiv.orgpaperAdded 30 Sep 2026, published 29 Sep 2026🧠 AI

STEPQuant is a spatial-temporal post-training quantization framework for the fixed-size recurrent states used by Delta-rule linear attention, which become a memory bottleneck under concurrent serving. It allocates bit precision based on error magnitude and memory lifetime (temporal) and fits per key-row/value-column scales based on state distribution and output impact (spatial), achieving near-FP32 accuracy at a 6-bit budget and beating uniform INT8 at 4-bit. Evaluated on Qwen3.8-27B and Kimi-Linear-48B-A3B-Instruct, and integrated into SGLang with custom GPU kernels, it delivers over 5x recurrent-state compression and up to 68.7% total serving memory reduction.

arxiv.orgpaperAdded 30 Sep 2026, published 29 Sep 2026🧠 AI

This paper proves that the Wasserstein distance between Laplacian spectra of original and LLM-reconstructed graphs is bracketed by two edge-count quantities, the net change in edge number and the symmetric difference, each scaled by 2/n, showing when these bounds coincide (pure addition or pure deletion of edges) and when they diverge (mixed editing, i.e. both inventing and losing edges). Using this theory, the authors analyze 135 graph reconstructions produced by three open-weight LLMs over 45 synthetic graphs, finding 77 one-sided outputs and 29 mixed outputs with a positive residual, revealing that aggregate distortion metrics mask qualitatively different editing policies across models, from copying the input to hallucination-heavy completion.

arxiv.orgpaperAdded 30 Sep 2026, published 29 Sep 2026🧠 AI

Researchers gave swarms of off-the-shelf coding agents a broad scope (from narrow topics to entire fields), literature access, and computing tools with a single standing instruction to make real progress and not stop, without supplying any scientific ideas themselves. Within weeks the agents produced research notes, paper-length drafts, and formal proofs (some verified in a proof assistant) across five areas of optimization theory and physical science, plus proposed untested lab experiments in a sixth, generating output faster than the human authors could review it. The paper argues research institutions are unprepared for this pace of capability growth and raises open questions about trust, credit, publication, and human oversight when review rather than production becomes the bottleneck.

arxiv.orgpaperAdded 29 Sep 2026, published 28 Sep 2026🧠 AI

This arXiv preprint introduces AI Night-Scientist, an agentic framework that uses GRPO reinforcement learning to counteract LLMs' low-entropy bias toward homogeneous outputs in open-ended scientific ideation. It models creativity along three axes (action, process, outcome) grounded in cognitive science, and trains models to depart from predictable reasoning via semantic guidance rather than just raising decoding temperature. Reported results show a 27.8% expansion in research-direction diversity, a 14.9% increase in contribution types, up to 32.0 points higher predicted citation impact, and 66.2 points higher originality versus the base model.

arxiv.orgpaperAdded 29 Sep 2026, published 28 Sep 2026🧠 AI

Latent Twin Operator (LTO) is a latent-space neural surrogate for time-evolving PDEs that uses a Convolutional Conditional Neural Process (ConvCNP) style encoder and decoder to ingest sensor observations at arbitrary, irregular placements and resolutions, paired with a learned latent evolution map that advances the encoded state directly between time points independently of the input and output discretizations. The paper derives an explicit convergence rate for context discretization error and validates it on a 2D heat-equation benchmark, showing LTO transfers across spatial resolutions with fixed parameters and, on Navier-Stokes, that direct latent evolution reduces error over long horizons compared to recursive rollout.

arxiv.orgpaperAdded 29 Sep 2026, published 28 Sep 2026🧠 AI

This arXiv paper proposes a certifier for structured, privileged agent actions such as paying an invoice, assembled from potentially corrupted documents and tool outputs, formalizing when multi-source corroboration can be trusted under indirect prompt injection. It uses a corruption-radius rule counted via minimum hitting sets over corruption-distinct evidence classes so laundered or re-published copies cannot fake a quorum, plus deterministic reconciliation and trusted anchors for single-sourced fields. Evaluated on 70,966 sanctions entities and 450 software-supply-chain provenance records, and against five LLM agents where a realistic injection fools all but one model and existing action-gating or provenance baselines are broken while the certifier admits no unsafe action.

arxiv.orgpaperAdded 2 Oct 2026, published 28 Sep 2026🛡️ Security

Speculative Uncertainty (SU) repurposes speculative decoding in reverse: a small open-weight draft model scores an agent's already-generated trajectory in a single forward pass, producing a failure-likelihood signal from output tokens alone without needing logits, weights, or repeated sampling. Phase-aware features separate reasoning and action spans and are calibrated against a verifiable objective to drive a pre-execution veto gate. On software engineering agents built on Qwen3-Coder-480B and closed-source Claude 3.5 Sonnet, the gate cuts execution error rate by 6-8 percentage points and token cost by 14-19%, transferring to out-of-distribution benchmarks without retraining.

arxiv.orgpaperAdded 26 Sep 2026🧠 AI

This preregistered study audits whether black-box LLM judges give stable, reproducible readings on shared serving endpoints, running two campaigns across 52,988 request attempts. Same-window repeat rankings only reached Spearman 0.400 against a required 0.90, and byte-identical next-day replays agreed at 0.78 against a required 0.99, showing that a model name on a shared endpoint is not a frozen measurement instrument. The authors trace the failure to label-to-meaning mapping bias, tiny candidate gaps relative to instrument noise, and non-deterministic outputs on identical inputs, and propose a three-level snapshot-identity ladder plus design rules and a reporting checklist for anyone freezing eval gates on LLM judges.

arxiv.orgpaperAdded 26 Sep 2026🧠 AI

CWE-Trace is a benchmark and diagnostic framework of 834 manually curated Linux kernel samples spanning 74 CWEs, built with a strict temporal train/test split to rule out data contamination, used to evaluate eight vanilla LLMs and 15 LoRA fine-tuned variants on vulnerability detection and CWE classification. The paper introduces two new metrics — Directional Failure Index (DFI) and Hierarchical Distance and Direction (HDD) — and finds that fine-tuning only recalibrates output thresholds rather than improving actual security reasoning, with the best model reaching just 52.1% detection accuracy (barely above chance) and under 1.3% Top-1 CWE classification accuracy.

arxiv.orgpaperAdded 26 Sep 2026🧠 AI

This paper introduces the "Superposition Linearity Hypothesis": when inputs from two distinct text streams are linearly combined, an LLM's output approximates a superposition (average) of the two individual next-token distributions. The authors show this linearity is an intrinsic property of the Transformer architecture that weakens during pretraining but can be restored via lightweight fine-tuning, and they propose a guided decoding method that exploits this to generate two coherent continuations from a single forward pass.

arxiv.orgpaperAdded 26 Sep 2026🧠 AI

This arXiv paper investigates subliminal learning, the phenomenon where a teacher model biased via a system prompt (e.g., toward malicious responses) transfers that hidden trait to a distilled student model through semantically clean training data like numeric sequences. The authors propose 'trait-direction drift' as the underlying mechanism — measurable preference gaps in teacher outputs induce trait-aligned weight updates during SFT — and introduce probe-space corridor regularization, a defense that constrains drift along a calibrated trait direction, reducing malicious-response transfer from 29.55% to 6.45% with minimal task-accuracy cost on Qwen models. It matters for AI security because it addresses a covert channel for injecting hidden malicious behavior into distilled/fine-tuned models without visible poisoned content.

arxiv.orgpaperAdded 26 Sep 2026🛡️ Security

This paper proposes an online safety monitor for deployed LLMs that converts a verifier model's signal into a real-time alarm via a threshold calibrated using risk control, flagging when the model's outputs can no longer be trusted as safe. Tested on mathematical reasoning and red-teaming datasets, the simple thresholded verifier approach matched more complex monitors built on sequential hypothesis testing.

arxiv.orgpaperAdded 26 Sep 2026🛡️ Security

This paper introduces CWE-Trace, a benchmark of 834 manually curated Linux kernel vulnerability samples spanning 74 CWEs with a strict temporal split to eliminate training data leakage. Evaluating eight vanilla LLMs including DeepSeek-R1 and fifteen LoRA fine tuned variants, the authors find contamination gives no real advantage and that fine tuning only shifts output calibration rather than underlying security reasoning, with the best model barely beating chance and CWE ranking accuracy staying very low. Two new diagnostic metrics are introduced to expose this calibration without comprehension failure mode.

arxiv.orgpaperAdded 26 Sep 2026🛡️ Security

PYPILINE is a malicious PyPI package detection system that builds a structured knowledge base of suspicious APIs from static analysis (AST + API call graphs) of known malicious packages, then uses an LLM agent workflow with RAG to retrieve that knowledge base and perform semantic analysis of new packages, outputting structured evaluation reports. The authors report 96.7% precision, 99.6% recall, and 98.1% F1 (5.7–21.6 points above baseline tools), with ~0.6s average detection time per package under 30 concurrent threads, plus an empirical study of common attack strategies and frequently abused APIs in PyPI malware.

arxiv.orgpaperAdded 26 Sep 2026🛡️ Security

ShellGames is an LLM-driven SSH shell simulator built for cyber deception and moving target defense, addressing prior LLM-honeypot weaknesses (no persistent state, hallucinated output, latency, subversion by curious attackers) via automatic chain-of-thought prompting, memory management for state coherency, speculative command execution, smart routing of complex commands to a sandbox, and subversion detection. The authors introduce a standardized benchmark spanning correctness, consistency, state tracking and robustness, reporting 0.898 command accuracy, 0.918 sequence-level consistency, 0.98 state-tracking accuracy and 0.95 robustness accuracy, all beating baselines by double-digit percentage points, and a 20-participant user study showing realism comparable to a real shell and better perceived coverage than traditional honeypots.

arxiv.orgpaperAdded 26 Sep 2026🛡️ Security

This arXiv paper introduces Knowledge Trap, a defense against model extraction attacks on commercial LLM APIs that lures attackers into querying a Honeypot Knowledge Graph (HKG) full of low-transferability knowledge via breadcrumb-guided exploration, wasting the attacker's query budget instead of blocking or perturbing outputs. Evaluated in medical and financial domains, it reduces surrogate model agreement by 6.2% on average while preserving accuracy for legitimate users, outperforming prior extraction defenses that degrade utility.

arxiv.orgpaperAdded 26 Sep 2026🛡️ Security

MCPTox is a benchmark for evaluating Tool Poisoning Attacks against LLM agents using the Model Context Protocol, where malicious instructions are hidden in a tool's metadata rather than its output. Built on 45 live real-world MCP servers and 353 authentic tools, it generates 1312 malicious test cases across 10 risk categories via three attack templates, and finds widespread vulnerability across 20 LLM agents: o1-mini has a 72.8% attack success rate, more capable models are often more susceptible due to stronger instruction-following, and even the best model (Claude-3.7-Sonnet) refuses less than 3% of attacks.

arxiv.orgpaperAdded 26 Sep 2026🛡️ Security

This arXiv paper addresses aligning the distribution of an attribute in a black-box generative AI model's outputs, such as fairness attributes or synthetic data representativeness, with a user-specified target distribution, using only repeated query access without model internals. The authors develop query-minimizing post-processing algorithms for exact and approximate alignment across multiple generated outputs, proving asymptotic optimality, and validate the approach on text-to-image generation and geocoded persona generation tasks.

arxiv.orgpaperAdded 28 Sep 2026, published 25 Sep 2026🧠 AI

This paper introduces READ (Read-only Expansion of Adapter Deltas), a method for composing multiple independently trained LoRA adapters into a single model without the weight-space interference that plagues naive merging. READ rewrites each adapter into a balanced canonical factorization and adds a one-directional coupling so new skills can read old adapters' input subspaces but never write into their output subspaces, requiring only a small trainable coupling row per new skill and folding cleanly into base weights with no inference-time routing cost. Evaluated across four benchmark suites and two model families, READ outperforms the strongest published composition baselines by over 20 points on SuperGLUE and over 7 points on a domain suite when adding skills sequentially.

arxiv.orgpaperAdded 28 Sep 2026, published 25 Sep 2026🧠 AI

GraphWrit3R is an end-to-end method for generating 3D scene graphs directly from 3D point clouds, Gaussian Splats, or both, outputting a structured JSON graph of objects, attributes, and relationships without relying on ground-truth object annotations or brittle multi-stage pipelines. It encodes point clouds via Sonata and Gaussian Splats via Chorus, fuses both modalities on a shared voxel grid with a novel per-voxel contrastive alignment loss, and decodes the result with a large language model, achieving state-of-the-art object class, predicate, and triplet recall on the 3DSSG benchmark while also supporting open-vocabulary querying.

arxiv.orgpaperAdded 28 Sep 2026, published 25 Sep 2026🧠 AI

FragToken is a training-time attack framework that manipulates an LLM to favor longer, non-canonical token sequences for the same output text, inflating autoregressive decoding steps and inference cost without proportionally lengthening visible responses. Naively maximizing token fragmentation degrades answer quality and breaks stealth, so the authors combine source-model self-distillation, capacity-aware filtering/budgeting, and BPE-Aligned Merging to keep utility largely intact; across four LLMs and three benchmarks it achieves a token inflation ratio (TIR) of 1.99–2.46 with only minor utility loss, framing this as a covert LLM supply-chain threat.

arxiv.orgpaperAdded 28 Sep 2026, published 25 Sep 2026🛡️ Security

This arXiv paper proposes checkability - whether a task has a cheap, deterministic intrinsic check that can reject incorrect outputs - as the deciding factor for whether local small language models (SLMs) can be trusted for network-automation tasks instead of sending sensitive configs, topologies and logs to a third-party frontier LLM. The authors build Touchstone, a local-first pipeline that runs seven off-the-shelf SLMs (1-8B params) to generate candidates, uses task-specific intrinsic checks to reject bad outputs, and escalates only unresolved cases to a frontier model, reaching 98.6 percent and 93.8 percent end-to-end accuracy on conflict-detection and intent-translation tasks while escalating just 16-17 percent of inputs. On TeleQnA, a knowledge-only benchmark lacking task-specific checks, Touchstone fails to match the frontier baseline, showing the escalation strategy only works when precise correctness checks exist.

arxiv.orgpaperAdded 28 Sep 2026, published 25 Sep 2026🧠 AI

This paper studies "decision models" like Jev, which map unstructured language directly to probability distributions over a fixed set of choices to drive routing, tool selection, and action triggering. The authors use an optimizer guided by the model's own option probabilities to craft short, fluent context additions that preserve the original question and gold answer yet flip the model's decision: 61.4% of initially correct decisions (312/508) are flipped for Jev, with 229 reaching ≥0.7 confidence in the wrong option, and three other decision systems show flip rates of 64.9%-73.2% across seven datasets. The results argue that treating these models' probability outputs as reliable decision interfaces is dangerous, since ordinary contextual phrasing — not adversarial-looking text — is enough to hijack them.

arxiv.orgpaperAdded 26 Sep 2026, published 24 Sep 2026🛡️ Security

This arXiv paper shows that dedicated decision models (e.g., Jev) which map language to probability distributions over finite choices can be "flipped" from a correct answer to a high-confidence wrong one by adding short, fluent, natural-seeming context — without altering the source, question, choices, or gold answer. Using an optimizer that refines context additions based on the model's option probabilities, the authors flip 61.4% of initially correct decisions in Jev within 64 accepted evaluations, and 64.9%-73.2% across three other decision systems on seven datasets. The findings raise concerns about relying on these models' probability outputs as trustworthy decision interfaces for routing, tool selection, or triggering actions.

arxiv.orgpaperAdded 26 Sep 2026, published 24 Sep 2026🧠 AI

PoEM predicts the outcome of RL post-training on a new reward function using models already trained on other rewards, without running additional RL. It shows that when a new reward is a linear combination of existing ones, the resulting log-policy is a corresponding linear combination of existing log-policies, and even for non-linear reward combinations, log-policies often span a low-rank subspace whose combination weights can be estimated from sample outputs alone. Validated on synthetic and real rewards across text and image modalities.

arxiv.orgpaperAdded 26 Sep 2026, published 24 Sep 2026🧠 AI

Minimally Invasive Steering Vector Optimization (MISVO) regularizes pre-logit activation steering using the local KL/Fisher geometry of the token distribution, letting a frozen model be adapted to a test-time reward without badly distorting its outputs. Across 1B-14B parameter models on preference and code-generation tasks, MISVO gets highest mean reward in six of seven settings while keeping diversity and coherence close to Best-of-N.

arxiv.orgpaperAdded 26 Sep 2026, published 24 Sep 2026🧠 AI

This paper identifies an "alignment illusion" in Multimodal Large Language Models: layer-wise visual-text similarity scores (CKA, SVCCA, MIR, leading principal-angle cosine) fail to consistently detect when visual tokens are corrupted with Gaussian noise, despite corruption sharply reducing task accuracy, across 13 MLLMs from 0.5B to 72B parameters spanning five model families. The authors trace this failure to anisotropic MLP down-projections in the shared language-model pathway that pull visual and text tokens toward common output directions (weight-induced alignment), and introduce the principal-angle gap (PA gap) as a metric that tracks task accuracy more reliably than existing scalar alignment scores.

arxiv.orgpaperAdded 26 Sep 2026, published 24 Sep 2026🧠 AI

This paper presents a distill-then-RL framework for query understanding (QU) in Roblox's game search, where an LLM is first supervised fine-tuned via teacher-student distillation to produce schema-compliant outputs (intent classification, query expansion, etc.), then optimized with reinforcement learning using component-specific rewards derived from live interaction with the search engine rather than a single end-to-end reward. On Roblox search, this component-specific RL raises NDCG@20 by 8.9 points over the SFT baseline and 3.5 points over end-to-end reward training, showing that decomposed, search-aware rewards better capture how each QU component affects downstream retrieval and ranking.

arxiv.orgpaperAdded 26 Sep 2026, published 24 Sep 2026🧠 AI

This paper describes a hypothesis-driven simulation workflow (using the Snowglobe simulator) for screening candidate LLM-based customer experience agents before production deployment, applied to Nubank's Card Delivery and Card Management chat-support agents in Brazil at 140M-user scale. Synthetic customers and simulated tool outputs let the team test multi-step agentic workflows without hitting production backends; across 4 deployed versions, simulated evaluator scores correlated strongly with production outcomes, and simulation-guided iteration lifted transactional NPS by 36.69 points and self-service rate by 8.82 points in live A/B tests.

arxiv.orgpaperAdded 26 Sep 2026, published 24 Sep 2026🧠 AI

A2M (Attraction-to-Manipulation) is a two-stage black-box framework for hijacking agents that use the Model Context Protocol to select tools from third-party MCP servers. The Attraction phase optimizes attacker-controlled tool metadata to raise invocation probability, while the Manipulation phase uses execution traces to craft adversarial tool outputs that steer agent behavior; on LiveMCPBench it achieves a 93.6% malicious tool invocation rate against GLM-4.6, inflates token cost 32.4x under a cognitive denial-of-service variant, and reaches a 74.4% mean attack success rate across information exfiltration, environment integrity compromise, and reasoning derailment, with partial transfer (63.6%/2.7x/24.5%) to four other models without re-optimization. The authors argue this exposes a semantic supply-chain risk in MCP tool ecosystems and call for stronger tool vetting and runtime isolation.

arxiv.orgpaperAdded 2 Oct 2026, published 22 Sep 2026🛡️ Security

This arXiv paper (cs.IT, submitted Sep 2026) reframes black-box LLM steganography as a sequential communication problem with causal feedback, introducing Burnashev Adaptive Posterior Matching (BAM) — a feedback-coding scheme combining posterior matching with a decode-and-confirm phase, backed by a cryptographic reduction proof for its security. Tested across three open-weight LLMs, BAM achieves 0-0.1% message error on 8-bit payloads in ~50 tokens (vs 10-17% for the strongest black-box baseline), and the authors demonstrate a full end-to-end covert communication protocol usable within ordinary-looking conversational text. This matters for AI security as it shows how agents/models could exchange hidden payloads inside seemingly benign LLM output, evading white-box watermark detection.

arxiv.orgpaperAdded 22 Sep 2026, published 21 Sep 2026🛡️ Security

This arXiv preprint defines and studies "Provider-Side Token Inflation Attacks" (PTIA), where LLM API providers covertly manipulate generation to produce longer outputs (and thus higher bills) while preserving task utility, testing five attack variants at the query, prompt, representation, and model levels that each inflate output length over 10x baseline. The authors identify a "PTIA saturation" effect tied to end-of-sequence token probability suppression, and build a lightweight single-probe black-box audit (no trusted reference model needed) that detects PTIA with 85.1% average detection rate and under 2% false-positive rate across four open-weight models, flagging 7 of 15 real-world LLM API services as PTIA-consistent.

arxiv.orgpaperAdded 22 Sep 2026, published 17 Sep 2026🧠 AI

This arXiv paper defines Provider-Side Token Inflation Attacks (PTIA), where a dishonest LLM API provider covertly manipulates generation at the query, prompt, representation, or model level to inflate billed output tokens while preserving apparent task quality. The authors instantiate five such attacks (each pushing mean output length past 10.2x baseline), identify a "PTIA saturation" effect tied to suppressed end-of-sequence token probability, and build a lightweight single-probe black-box audit that needs no trusted reference model, achieving 85.1% detection with under 2% false positives across four open-weight models — and flags 7 of 15 real-world LLM API services as PTIA-consistent.

arxiv.orgpaperAdded 18 Sep 2026, published 17 Sep 2026🛡️ Security

Diffusion TV is an interactive AI art installation built from a modified CRT TV that lets audiences control the clarity of AI-generated images and sounds by physically manipulating the antenna, metaphorically enacting the denoising process of diffusion models. A tuning knob switches between three channels of AI-generated animals (extinct, endangered, speculative), embedding the interaction in a temporal and ecological narrative. The work argues for embodied, tangible interaction as an alternative mode of explainable AI that foregrounds the generative process itself rather than final outputs.

arxiv.orgpaperAdded 22 Sep 2026, published 4 Sep 2026🧠 AI

A preregistered audit of whether identical requests to the same-named LLM API return stable outputs over time, finding they largely do not: across 52,988 audited requests, same-window repeat rankings hit Spearman 0.400 against a required 0.90, and next-day byte-identical replays hit 0.78 against a required 0.99. The authors trace the failure to label-mapping bias, candidate gaps far below the instrument's noise floor, and non-deterministic outputs for identical inputs, tested across four providers and self-hosted batch-invariant kernels, concluding a model name on shared infrastructure is not a frozen measurement instrument and that evaluations must characterize instrument noise before freezing any gate on it.

arxiv.orgpaperAdded 22 Sep 2026, published 3 Sep 2026🧠 AI

GRADSOLVE is an open-source JAX library for solving and reverse-mode differentiating low-dimensional ODE ensembles on NVIDIA GPUs. It records the steps an adaptive solver accepts and differentiates a fixed-step replay, producing the exact discrete adjoint (matching Diffrax's default output) more cheaply than an adaptive loop, running 2.8x faster as a forward solver and 5.6-14.1x faster for gradients than Diffrax's checkpointed adjoint at matched accuracy.

arxiv.orgpaperAdded 22 Sep 2026, published 2 Sep 2026🧠 AI

This arxiv paper introduces the TrustX Agent Risk Classification Framework (ARC), a structured instrument for risk-tiering internally built agentic AI systems across seven system types. It combines a twelve-dimension scoring rubric with a GPA + IAT classification model and a five-level autonomy framework to produce a three-tier governance output with mapped controls, including a Coding Assistant extension, and provides an interactive tool plus a worked example for AI risk officers and governance practitioners.

arxiv.orgpaperPublished and added 18 Jul 2026🛡️ Security