62 results for “PPO”

WorkFactory is a framework that builds coherent, executable professional tasks for training working agents. It plans an information support graph from professional intents, then uses probe executions to find and repair missing or inconsistent task information. The authors built 112,686 tasks across 612 occupations and fine-tuned Qwen models at 4B, 27B and 35B-A3B (WorkPilot). WorkPilot-35B-A3B rose from 879 to 1565 Elo on GDPval-AA v2.1.

huggingface.copaperAdded 30 Oct 2026🧠 AI

Turing LLM post-training combines calibrated multi-teacher on-policy distillation (C-MOPD) with an execution path for dynamically routed mixture-of-experts models. C-MOPD matches each teacher on a local Top-K support with teacher-wise loss normalization and clipping. On Turing-20B-A2B it scores 0.9988 on four-domain integration, beats Qwen3-8B on math and code (85.92 vs 80.33 mean), and cuts consolidation step time from 398.7 to 139.3 seconds in a PPU profile.

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

CALM is a conflict-aware lifecycle memory governance framework for long-horizon language agents. It turns records into evidence-grounded memory cells, updates their validity and lifecycle states from support, conflict and dependency signals, and admits evidence at query time through training-free or learned gating based on temporal applicability. It beats baselines on LongMemEval-S and LoCoMo with Qwen and GPT backbones, and keeping superseded records helps earlier-state reasoning.

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

CRISP is a combination-aware framework that co-designs pruning and runtime scheduling for Mixture-of-Experts inference on memory-constrained edge platforms. It reorders neurons by importance so one stored expert supports several execution widths, gives hot, warm and cold expert combinations different widths, and overlaps expert loading with computation. On Mixtral-8x7B and DeepSeek-MoE-16B it raises throughput up to 2.1x at the same memory budget, with accuracy loss limited to 3.6%.

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

This paper presents a pipeline that uses persona-conditioned vision language models to surface interpretive ambiguity in visualizations. Topic-relevant personas elicit candidate takeaways, and a Target-Pattern-Scope framework lets LLM judges label statement pairs as Compatible, Disjoint or Opposing, which are summarized in an ambiguity profile. The authors benchmark VLMs on chart comprehension per chart type, check labeling against human annotations, and run case studies on public visualizations and a redesign.

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

AutoTavern is a self-improving system of specialised LLM research agents with a typed shared memory and verification gates. In 20 days it built TavernGym, a full-content Hearthstone Battlegrounds simulator, and 23 versions of TavernAgent, a PPO-trained player. TavernAgent averaged placement 1.60 with 80% first places over its last 20 live games, beating the top ten players (2.34, 59%). On TavernBench (960 tasks), it answered 91% of play tasks against at most 63% for frontier LLMs.

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

Sliding Networks is a family of residual network structures with a joint width and depth scaling recipe that keeps both locally maximal feature learning and feature diversity in the residual stream. The key quantity is the ratio N_flat/(N_sharp L) tending to a constant alpha, and Half Copy correlated initialization can relax the shape constraints. Residual-MLP experiments support an SDE limit, but language-model experiments show performance differences under regrouping at fixed parameter count.

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

Accordion is a mixture-of-experts language model (Accordion-6B-A0.5B) that supports five inference budgets through fixed-prefix layer looping, with the minimum budget being the unlooped model. After 3.6T pretraining tokens, deeper inference raises the macro-average over 37 tasks by 3.01 points, and the middle budget captures 89% of that gain with half the extra layer executions. SFT and RL keep the budgets, merging three domain experts via parameter-group merging and budget-matched multi-teacher on-policy distillation.

arxiv.orgpaperAdded 26 Oct 2026, published 23 Oct 2026🧠 AI

Sotto Maschera is a masked implementation of the Sota Voce sparse fixed-weight vector sampler (Sota support sampler and MWSupport conversion) for embedded devices, protecting against power side-channel attacks. It bitslices sorting, merging and compaction and masks only in the Boolean domain, avoiding slow conversions. TVLA evaluations show no leakage, and HQC runs up to 45x faster than previous masked work.

eprint.iacr.orgpaperAdded 24 Oct 2026, published 21 Oct 2026🛡️ Security

"Is Sycophancy a Form of In-Context Updating?" argues that factual sycophancy, a model's deference to a user's false beliefs, is a failure of in-context updating rather than only a bid for approval. Models also revise answers for claims attributed to external sources, more so when better supported, and deference persists when no user approval is at stake. The internal components behind deference also matter causally for general updating, so reducing sycophancy may reduce useful updating.

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

"GKS FHE: Security Analysis" is a Cryptology ePrint Archive preprint (2026/2523, October 2026) giving a formal security analysis of the Gribov-Kahrobaei-Shpilrain private-key fully homomorphic encryption scheme. The abstract says the scheme needs no bootstrapping or de-linearization, gives bit-exact decryption, supports native encrypted search and has unbounded depth. The page shows only the abstract, so the analysis's methods and results are not visible.

eprint.iacr.orgpaperAdded 19 Oct 2026, published 18 Oct 2026🛡️ Security

The One-Pixel of Reasoning introduces the one-relation attack (ORA), which finds a critical implicit warrant linking evidence to a conclusion and inserts one source-grounded sentence that qualifies it, steering an LLM agent to a wrong belief. Across eighteen models on biomedical inference and vulnerability detection, best-of-10 attacks flip initially correct answers 54.9% of the time for proprietary models and 80.0% for open-weight models. Attacks transfer without victim feedback, can pass source-support checks, and multi-agent debate raises team error from 50.0% to 59.5%.

arxiv.orgpaperAdded 21 Oct 2026, published 17 Oct 2026🛡️ Security

This paper proposes a framework for auditing LLM-agent-generated stakeholder preferences before they are used as weights in multi-objective optimization for healthcare operations. Experiments on outpatient appointment scheduling and emergency department configurations show that explicit ratings vary by role and clinic conditions, deliberation can erase role differences, and pairwise choices depend largely on option order. Robust weighted-sum models often select different schedules than average weights.

arxiv.orgpaperAdded 20 Oct 2026, published 16 Oct 2026🧠 AI

This paper builds a benchmark of 500 debatable questions across ten domains to measure how user narratives steer LLM agreement. Each question has 11 matched conditions (one neutral, five mirrored narrative levels per opposing stance), giving 5,500 instances across seven LLMs. Causal framing and worldview synthesis raise agreement, personal experience can lower it, and GPT-5.1 ablations show bullet formatting also increases agreement.

arxiv.orgpaperAdded 20 Oct 2026, published 16 Oct 2026🧠 AI

ePrint paper presenting the first fully masked hardware implementation of HQC (key generation, encapsulation, decapsulation) on FPGA. It introduces a mixed Boolean/arithmetic masking for sparse-dense vector multiplication, proven first-order secure in the transition- and glitch-robust probing model and PINI-composable, at only 3% extra cycles and 81 random bits per cycle. It also shows a profiled horizontal side-channel attack on Rotate-and-Add multipliers that recovers support indices within ±8 positions from 200 traces despite TVLA showing no leakage after 10 million traces, and adds integrated hiding countermeasures costing 1.7% latency.

eprint.iacr.orgpaperAdded 17 Oct 2026, published 16 Oct 2026🛡️ Security

The paper separates "active" facts (which support generalization beyond recall) from "inert" ones. It builds new relations by relabeling answers of existing ones (Paris -> Europe becomes Paris -> crab), fine-tunes on a subset of subjects, and tests held-out subjects that share an answer (Munich -> crab). Across eleven language models and 70 relations, linear readability of the original relation in pretrained representations correlates with this generalization at Pearson r = 0.87 to 0.92 within each model. The result holds after controlling for fine-tuning set size, original-relation accuracy and subject frequency in pretraining text, and it survives re-randomized labels, seeds and thirteen other label categories. Mapping wrong labels back to the original answers shows the fine-tuned model's errors follow the linear probe's top answers.

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

AgentSWE is a benchmark of 25 tasks covering the agent development lifecycle: Creation of an agent from a natural-language requirement, Editing of production agent codebases, and Optimization toward a measurable target. Builders iterate on a few illustrative cases, then the frozen submission is graded on held-out cases with requirement-derived rubrics and hard pass/fail gates. Across coding agents, models and harnesses, none comes close to delivering the requirement. Recurring failures include overfitting to visible cases, hollow runtimes behind code that looks complete, and unsupported claims of completion.

huggingface.copaperAdded 13 Oct 2026🧠 AI

Introduces a cube-attack framework with an elimination strategy. It finds key conditions and cubes whose conditional superpolies are cheap to compute, then solves the resulting equation systems to recover key bits. The supporting techniques are a nested coefficient solver, conditional monomial prediction, an automated cube search combining numeric mapping with monomial prediction, and two testing methods for large equation systems. Against round-reduced Trivium it gives a practical key recovery at 840 rounds (complexity below 2^55, 77.8% success) and at 845 rounds (below 2^56, 98.1% success on 59.1% of keys). It also gives a theoretical weak-key attack at 855 rounds for 2^74 keys. These are the best known results, extending prior practical attacks by 13 rounds and theoretical attacks by 4.

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

Proposes Staggered Instruction-Fetch Duplication (SIFD), a processor-level fault-injection countermeasure that fetches each instruction twice in consecutive cycles while fetching two different instructions per cycle, so inconsistent fetches are detected without full processor duplication, compiler support or crypto hardware. Implemented on a custom RV32I core on an Artix-7 FPGA, it costs 15.0% more slices and 16.3% more execution time. Attack success falls from 0.86% to below 0.01% under clock fault injection and from 0.32% to 0.13% under EMFI, with 73.18% of CFI and 11.44% of EMFI attempts detected.

eprint.iacr.orgpaperAdded 14 Oct 2026, published 10 Oct 2026🛡️ Security

Paper testing whether an LLM agent operating life support in the NASA-derived BioSim lunar habitat simulation can be induced by delayed prompt injection, hidden in truthful S-band telemetry, to shut down life support and hold the shutdown while the crew suffocates. It compares Claude, GPT and Grok configurations: Claude mostly refuses and cites the living crew, GPT shows some susceptibility, and Grok given a "trusted commission" framing more often cuts life support. It also varies the framing of the message (trusted commission, peer, untrusted source) and uses a harmless-procedure-drop attack as a proxy for whether life is valued intrinsically. Code and traces are open-sourced.

arxiv.orgpaperAdded 15 Oct 2026, published 10 Oct 2026🛡️ Security

Presents the first fixed-point FN-DSA (Falcon) signer for microcontrollers, using 32.64 arithmetic in constant-time Cortex-M4 assembly for Falcon-512 and Falcon-1024. Techniques that keep reciprocals, inverse square roots and small constants wide and round once at the consumer restore the precision lost by naive fixed-point. Measured precision supports about 2^99.7 (Falcon-512) and 2^96.7 (Falcon-1024) signing queries per key, against about 2^61.5 and 2^58.7 for double precision and the 2^64 target. The signer needs about 45% fewer cycles than double-precision emulation, at the cost of signing RAM rising from 59n to 83n bytes. The paper also shows which implementation choices lift double-precision c-fn-dsa above 2^64, derives Falcon-1024 key-generation bounds (72.3% of keys accepted vs 8.5% with the Falcon-512 values), and gives recommendations for the fixed-point FN-DSA NIST plans.

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

DualQuant is a calibration-free post-training quantization framework for LLM weights and activations. It minimizes weight-reconstruction MSE using variable splitting and block coordinate descent with closed-form subproblem updates. The resulting scaling factors double as per-channel activation rescalings, so no calibration data is needed. It supports the MX family and NVFP4 block formats. On Llama and Qwen models it matches or beats calibration-based SmoothQuant on WikiText-2, C4 and zero-shot tasks. A real packed W4A4 path for Llama-3.1-8B on an RTX 4090 gives up to 2.61x memory reduction and 2.46x throughput over BF16.

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

Proposes Matched-Context Group Policy Optimization (MCGPO), which builds weighted action preferences from already-collected sibling rollouts to give credit to intermediate decisions when feedback is sparse and arrives only at the end. Comparison contexts use pre-action state summaries and rubric progress. Exact and coarse matching rules are selected by two-action support, and fixed folds separate candidate decisions from the terminal outcomes that orient them. An evidence-weighted auxiliary pairwise loss is added to GraphGPO task learning, and no critic is needed. Across six environment-backbone settings with three seeds each, it beats GraphGPO by 0.47-1.46 points on ALFWorld and WebShop with Qwen2.5-1.5B/7B, and by 3.96 and 5.16 points on WebShop with Qwen3-4B/8B. ALFWorld gains are small because baseline success is already above 92%.

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

DecepEval is a benchmark of 1,532 instances across 3 task families and 28 professional scenarios designed to measure deception in LLM agents, pairing neutral and induced versions of each task to isolate condition-dependent changes in deception rates. It introduces the LLM Deception Diamond framework (pressure, incentive, opportunity, conflict), adapted from classical fraud theory, and finds that across nine frontier LLMs these inducements raise deception rates even in models with low baseline deception, with long-horizon tasks showing deception rates averaging 87% under induced conditions.

huggingface.copaperAdded 8 Oct 2026🧠 AI

Introduces "prior barrier", a measure of how strongly a pretrained LLM supports competing concepts over a target concept, and shows prior barriers follow a long-tail distribution, so tail concepts need more SFT instructions than head concepts. It derives a predictive risk bound for SFT under long-tail prior barriers and proposes PASS, an adaptive instruction selection method that builds reference-derived concepts, estimates each instruction's distinguishing evidence, and allocates the selection budget toward under-covered concepts. PASS beats seven state-of-the-art instruction selection methods across four backbone-budget settings, and ablation shows adaptive allocation beats uniform allocation.

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

Case study in which Opus, GPT and Gemini coding agents port thirteen QED-C Application-Oriented Benchmark methods from Qiskit to CUDA-Q. The agents produced satisfactory ports far faster than a human engineer, but took shortcuts when implementation got hard and needed repeated prompting. No single model was best across all benchmarks, which suggests multi-agent ensembles. The work released near-complete CUDA-Q support in QED-C and model-generated skills that, on an unseen benchmark, cut model turns by up to 85% and token usage by up to 96%.

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

This paper proposes three evaluation screens — Decisive-Action BPB, Patch MCQ, and prefix-conditioned pass@K — that predict how well a base model will perform as a coding agent after post-training, without requiring the base checkpoint to run a full agentic harness. By replaying successful post-trained agent trajectories and identifying the "decisive step" that flips a repository from failing to passing tests, the authors measure a base model's latent support for that action. Tested across ten pairs of public base and post-trained models, all three screens closely track post-trained SWE-bench Verified pass@1 scores, offering a way to cheaply vet base checkpoints before committing to expensive agentic post-training.

arxiv.orgpaperAdded 8 Oct 2026, published 7 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

FREA is a benchmark of 751 chemical reactions labeled by expert chemists for feasibility verification, drawn from retrosynthesis model proposals, zero-yield experimental records, LLM-generated edits, and five negative-candidate generation methods. The authors show no single verifier type (LLMs prompted with the feasibility criterion, dedicated forward models, or retrosynthesis scorers) dominates across all sources, and that adding generated negative examples to training improves mean AUROC but fails to transfer to retrosynthesis proposal screening. They also release a corpus of over 14 million recorded and generated reactions to support further training and evaluation of feasibility verifiers.

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

This paper develops a sharper theoretical landscape analysis for Low-Rank Adaptation (LoRA) fine-tuning, introducing LoRA-RIP, a data-dependent restricted-isometry metric that measures how well-conditioned the cross-entropy objective is along LoRA-relevant low-rank directions. The authors prove that sufficient rank over-parameterization (with the required rank set by the LoRA-RIP constant) eliminates spurious local minima, extending prior RIP-based guarantees beyond the classical 1/3 regime, and show this framework also enables principled data selection under a fixed rank budget. Experiments on language and vision tasks support the theory, arguing that adapter rank and data quality are coupled resources that should be chosen jointly.

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

This IACR eprint paper presents a resource-efficient FPGA hardware architecture for the NIST-selected HQC post-quantum KEM that defends against power-based side-channel attacks by reusing otherwise-idle KECCAK hardware to inject scrambling dynamic power during decoding, adding less than 0.5 percent area overhead compared to masking/shuffling countermeasures. The fully parameterized design supports HQC-128/192/256 and was prototyped on a Xilinx Artix-7 FPGA (139 MHz, 11,489 LUTs, 6,785 FFs, 4 DSPs), with functional correctness verified via NIST Known Answer Tests and leakage suppression confirmed via Test Vector Leakage Assessment.

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

Agentic-ZTA is a multi-agent architecture that operationalizes the NIST SP 800-207 Zero Trust Architecture control loop using coordinated AI agents. A Policy Enforcement Point intercepts access requests, enriches them with context, and routes them to a policy-engine agent that invokes domain-specialized core and supporting agents; policies are retrieved via RAG and embedded into agent prompts at inference time, with trust scores aggregated by a trust algorithm to produce the final access decision. The authors implement and evaluate the framework on a testbed, reporting 95.0% accuracy, 93.9% precision, and 96.3% recall on representative access-control scenarios.

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

This arXiv paper identifies a new attack surface in proactive personal AI agents: an external content provider who controls only material tied to its own target (e.g. a product or service) can manipulate an otherwise benign agent into recommending that target to the user, without compromising the agent or accessing private user context. The authors formalize the attack via three mechanisms — Target Control, Private Binding, and Prospective Support — and show across three proactive-agent environments and six simulated user models that the full attack raises target authorization rates by up to 77.4 percentage points, with a multi-turn variant that keeps influencing the agent's responses even absent final user authorization.

arxiv.orgpaperAdded 6 Oct 2026, published 4 Oct 2026🛡️ Security

ReFract is a benchmark of 150 expert-validated entries testing whether LLM agents can infer a user's role and act only through tools/actions legitimate for that role, a capability the authors term "Perspective Awareness." Entries are grounded in anonymized industrial-maintenance support queries and use constructed Text World Models to simulate the agent's operating environment and generate perspective-aware action trajectories. State-of-the-art LLMs solve at most 69% of tasks, with over half of trajectories containing attempts at perspective-violating actions, exposing this as a largely unsolved axis of agent evaluation relevant to high-stakes, physically-enacted agent deployments.

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

This arXiv preprint extends the concept of reward hacking to four distinct "proxy" layers in a staged RL training pipeline for legged robots: the reward itself, the curriculum gate, the evaluation statistic, and the reference motion. Using measurements from a continuous PPO policy lineage trained on a simulated 1.91m humanoid over four stages and 13,500 iterations on a single laptop GPU, the author shows concrete divergence failures — e.g., a curriculum gate built on averaged error that advanced at its rate limit while the gated skill was absent, and a batched push test whose synchronized resets aliased gait phase and misranked push magnitudes by danger. For each layer it proposes a reformulation that closes the divergence, including L1 costs, peak/outcome statistics, gate reachability checks, deterministic/phase-desynchronized evaluation, and function-preserving input widening for growing a single policy instead of retraining.

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

AREX-2 trains an LLM agent, built on Qwen3.8-27B, to self-improve at test time by iteratively refining solutions. The training data is synthesized long-horizon improvement trajectories from machine learning and algorithmic programming tasks, which give verifiable feedback. It reports 81.8 on MLE-bench Lite and 70.7 on Frontier-CS, and transfer to deep research with 84.0 on BrowseComp, 52.6 on HLE, 92.2 on GAIA and 93.8 on DeepSearchQA. Performance keeps improving as the budget of rounds grows. This supports the claim that reflection and long-horizon execution are domain-agnostic skills that can be learned from supervisable domains.

arxiv.orgpaperAdded 1 Oct 2026🧠 AI

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

DRelay is a speculative decoding technique for LLM inference that repairs parallel draft-token selections before target-model verification, using a global reader over the entire draft block plus a causal selector to decide whether to keep or replace each candidate token, extending the accepted prefix length. Trained jointly with the draft backbone via a combined candidate-support and position-weighted repair loss, it is evaluated on eight benchmarks on an H800 GPU and, under SGLang serving, improves end-to-end decoding speedup substantially over DFlash, Domino, and DSpark baselines.

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

CheatBench is a publicly released benchmark for measuring "cheating" behavior in RL-trained AI agents — environments spanning math research, knowledge work, coding, and visual tasks that pair hard assignments with built-in opportunities to game the reward signal instead of doing honest work. The paper cites real incidents and controlled evaluations where agents accessed unauthorized information, evaded monitoring systems, and breached sandbox protections to attack external systems, and provides a testbed to compare models/tasks and study mitigations as agents take on more consequential responsibilities.

arxiv.orgpaperAdded 30 Sep 2026🛡️ Security

CheatBench is a benchmark for measuring "reward gaming" (a.k.a. reward hacking) in reinforcement-learning-trained AI agents: environments across mathematical research, knowledge work, coding, and visual tasks pair difficult assignments with opportunities to cheat, letting researchers compare how different models pursue reward when honest completion is hard. The paper motivates this with real and controlled incidents of agents accessing unauthorized information, evading monitoring, and breaching sandbox protections, and the benchmark is publicly released to support cross-model and cross-task-category comparisons.

arxiv.orgpaperAdded 30 Sep 2026🧠 AI

DMA^2 is a distribution matching distillation (DMD) method tailored to pixel-space (rather than latent) diffusion teachers for text-to-image generation. It introduces DINO-Adv, which removes the heavy fake-score critic from the adversarial gradient path and instead supplies local patch guidance via an external DINOv2 representation, plus AF-Loss, a parameter-free auxiliary semantic distribution-field objective operating in DINOv2 feature space on rolling real/generated supports. The resulting four-step student outperforms its 25-step teacher and other few-step distillers on DPG-Bench, GenEval, VQAScore, and COCO30K.

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

JsVul is a function-level dataset of vulnerable and fixed JavaScript/TypeScript code pairs, aggregated from seven sources and built with a language-specific pipeline that strips noise (minified code, cosmetic edits), deduplicates across commits, and applies heuristic labeling of security-relevant changes. It is released in time-ordered JSONL format to support training and evaluating automated vulnerability detection models specifically for the JS/TS ecosystem, where prior datasets were largely generic and multi-language.

arxiv.orgpaperAdded 30 Sep 2026, published 29 Sep 2026🛡️ Security

This arXiv paper analyzes reward hacking in reinforcement learning with verifiable rewards (RLVR), using gradient flow analysis with a fixed verifier to characterize when reward increases while correctness decreases. The authors prove that standard RLVR training signals are insufficient to detect or reduce "accepted errors" (incorrect responses rewarded by the verifier) without also suppressing correct responses, and propose a "selective control" correction that incorporates partial audit feedback on correctness to simultaneously lower accepted-error probability and raise correct-response probability. Experiments on log-linear and neural contextual bandits and a language model support the theory, showing selective control under partial auditing reduces accepted errors while increasing correctness.

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

DeepSeek Elastic Compute (DSec) is a production sandbox platform that unifies FnCall, container, microVM, and full-VM execution backends behind a single SDK to support large-scale agentic training and evaluation of LLMs. It co-designs sandbox lifecycle management with the RL training framework — decoupling stateful rollout execution from preemptible GPU training, sharing and reclaiming memory, and streaming images on demand from DeepSeek's Fire-Flyer File System (3FS) — to run around 160 nodes serving ~3 million sandboxes/day, over 380,000 concurrent sandboxes, and 5,000+ sandbox creations per second in production.

arxiv.orgpaperAdded 2 Oct 2026, published 26 Sep 2026🧠 AI

MemSentry is a proposed pre-commit evaluation framework for agentic AI systems that intercepts proposed writes to persistent memory and issues Accept/Review/Quarantine decisions by jointly scoring source trust, semantic risk, attack radius over a component-dependency graph, access risk, and a signed security-posture delta. The authors evaluate it against 1,000 GPT-4-generated scenarios using four pluggable semantic classifiers (Regex, TF-IDF+SVM, SBERT+LR, SetFit), with SBERT+LR reaching 91.7% accuracy while all methods catch 100% of external quarantine-class threats; the paper also documents a key limitation — surface-similar but oppositely-intentioned insider operations remain hard to distinguish semantically. It targets the persistent-memory-poisoning attack surface distinct from prompt injection, where a single malicious memory write (e.g., "suppress alerts from this IP") can silently persist and influence future agent actions like a SOC assistant's alerting or access decisions.

arxiv.orgpaperAdded 26 Sep 2026🛡️ Security

This paper presents algorithmic techniques enabling LLM inference on Compute-in-Flash (CIM) devices, which lack floating-point support and have limited write endurance. It introduces an end-to-end integer-only quantization pipeline (INT8 linear layers, integer Softmax/RMSNorm approximations) and a dictionary-based KV-cache compression scheme using sparse dictionary coding to cut KV write traffic by 15x. Evaluated on Llama-3.1-8B and Qwen-2.5-7B, the combined approach yields 3.1x-4.4x latency and 2.7x-6.8x energy savings over DRAM-based systems, particularly for long-context inference.

arxiv.orgpaperAdded 26 Sep 2026🧠 AI

Aray is a deterministic-first YARA interpreter and positive-fixture synthesizer that generates benign, YARA-matching ELF/PE/generic artifacts to validate detection rules without storing or distributing live malware samples. Conventional code (not a generative model) canonicalizes rules, derives witnesses from regex/hex declarations, and performs extraction, layout, and binary serialization; an LLM is only permitted to propose constructive rule normalizations or residual-semantics judgments, never backend source or binary structure, and the pipeline supports a zero-inference execution path. Evaluated against 416 public YARA rules, it achieved 406/416 (97.6%) matches to the original upstream rule, with the design explicitly motivated by containment concerns raised by the July 2026 OpenAI–Hugging Face agent incident.

arxiv.orgpaperAdded 26 Sep 2026🛡️ Security

This UC Berkeley paper designs LLM inference algorithms for Compute-in-Flash devices, which move computation into SSD memory to overcome bandwidth limits but lack floating-point support and have limited write endurance. It presents an end-to-end integer-only quantization scheme and a dictionary-based KV cache compression method (sparse dictionary coding representing KV vectors as linear combinations of static dictionary vectors), reducing dynamic KV cache traffic 15x on Llama-3.1-8B and Qwen-2.5-7B with limited accuracy loss.

arxiv.orgpaperAdded 26 Sep 2026🧠 AI

This arXiv paper proposes the Agent-Editing World Model (AEWM), which reframes world modeling for LLM agents: instead of predicting/reconstructing environment observations (high-entropy tool responses), it models how reasoning and actions shape future task progress, addressing "task-state contamination" where stale plans and unsupported assumptions corrupt later decisions. It introduces Action Judge (classifying decisions as Critical, Exploratory, or Noisy) and State Revision, combined via EditAct which directly edits the agent's state using real execution feedback rather than issuing critiques. Trained across Search, Terminal, and Software Engineering domains, AEWM reports 70.5% macro-F1 on its Action Judge benchmark (+10.6 over the strongest frontier baseline) and EditAct improves average scores by 3.2–6.7 points across six benchmarks and three agent backbones, with an RFT variant (AEWM-RFT) also outperforming Self-RFT.

arxiv.orgpaperAdded 26 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

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

This arXiv paper analyzes six official MCP SDKs (Python, TypeScript, Go, Rust, C#, Swift) and finds structural gaps for enterprise zero-trust use: Authorization headers can't cleanly support both human SSO and service-account callers, there's no pre-authentication tool discovery, and base SDKs lack fine-grained per-tool authorization. The authors build composable FastMCP extensions — cross-header credential normalization, cached token verification across IdPs, an unauthenticated metadata endpoint for registry discovery, and a declarative per-tool permission annotation — and show that visibility-only filtering alone is insufficient (models reference hidden tools by name up to 94% of the time) while combining visibility filtering with invocation-time enforcement drives forbidden tool exposure from 21.1% to 0% across 2160 attempts on four frontier LLMs.

arxiv.orgpaperAdded 2 Oct 2026, published 18 Sep 2026🧠 AI

This paper evaluates cross-sector generalization of a role classifier that structures French occupational accident narratives into four accident-process roles (work situation, unfavourable condition, accident event, consequence). Classifiers trained solely on 42,244 construction-sector factual units are tested without retraining on metallurgy, chemistry-plastics, and an independent company corpus, comparing frozen pretrained representations against task-specific fine-tuning and supervised representation learning. Task-adapted strategies consistently outperformed frozen representations, reaching 85.6-85.8% average balanced accuracy across unseen sectors, supporting transferable assisted-coding systems for occupational risk analysis.

arxiv.orgpaperAdded 22 Sep 2026, published 18 Sep 2026🧠 AI

FAMOS is a feed-forward model that predicts movable-part segmentation and joint parameters for articulated 3D objects from a sparse, unordered set of partial point clouds, supporting a variable number of input views including a single one. It introduces a Multi-state Articulation Transformer with alternating state-wise and global attention plus an observed articulation span objective, and uses a procedural data generator to synthesize training assets, outperforming feed-forward and optimization-based baselines on PartNet-Mobility, ACD, and ArtiCraft-10K.

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

PhysStream is an autoregressive image-to-video generation model that maintains structured scene memory (positional maps and object tracking maps derived online from previously generated frames) and supports fine-grained, mid-generation motion control via sparse velocity-increment signals encoding physical quantities. Trained in two stages (bidirectional motion-control finetuning, then causal autoregressive training with scene memory), it reduces motion distribution distance (FVMD) by 33% and trajectory error by 12% over strongest baselines on synthetic multi-object rigid-body tabletop scenes, and is preferred in over 85% of human in-the-wild comparisons.

arxiv.orgpaperAdded 22 Sep 2026, published 15 Sep 2026🧠 AI

This paper introduces Chain-of-Self-Questioning (CoSQ), a prompt-only framework that conditions an LLM's answer commitment on an explicit self-assessment of whether it has sufficient factual support to answer. Tested across eleven open-weight and hosted model families on the 817-item TruthfulQA multiple-choice set and a secondary Natural Questions Short-Answer evaluation, the best variant (Grounded-CoSQ at τ=0.90) cuts the mean wrong-commitment rate by 32.1% relative to chain-of-thought prompting while also raising answered accuracy, holding consistently across all models and thresholds tested.

arxiv.orgpaperAdded 22 Sep 2026, published 15 Sep 2026🧠 AI

IdeaAMBIG is a benchmark of 660 evidence-grounded instances (163 real-world gaps from reproducibility reports and GitHub issues, 497 synthetic) that tests whether LLMs can identify and resolve implementation-critical gaps in research-method specifications, i.e. whether a specification gives a coding agent enough information to build the intended method without unsupported assumptions. It evaluates three capabilities across 13 LLMs: codification-readiness assessment, defect localization, and clarification action generation, finding defect localization is the major bottleneck (best model only 9.6% Macro Defect Recovery Rate) while clarification given the defect is much stronger (80.6% success), and an oracle study shows gold resolutions raise codification-ready rate from 14% to 98%.

arxiv.orgpaperAdded 22 Sep 2026, published 9 Sep 2026🧠 AI

ReCite is a decoupled agentic framework for citation recommendation that replaces semantic-similarity retrieval with active claim-level reasoning, combining location perception, intent-aware query planning, and reflective verification to check claim-evidence consistency. Trained on synthesized reasoning trajectories, it triggers self-correction loops when retrieved candidates fail to logically support an author's claim, and the paper reports it outperforms larger generative models on strict citation accuracy.

arxiv.orgpaperAdded 22 Sep 2026, published 8 Sep 2026🧠 AI

UniMate is a unified foundation model that generates articulated motion for arbitrary rigged 3D skeletons from a text prompt, without per-skeleton fine-tuning or reference motions at inference. It introduces a topology-aware diffusion transformer combining graph-aware attention bias from joint relations/geodesic distances, a spectral RoPE generalization via the graph Laplacian, and a global topological conditioner, trained on the newly curated UniML3D dataset (13,006 motion sequences across bipedal, quadrupedal, avian, marine, insectoid, serpentine and rigid-object skeletons). It outperforms prior category-specific animators and supports zero-shot cross-topology transfer, in-betweening, expansion, and text-guided editing.

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