Don't Lose the Thread: Empowering Long-Horizon LLM Agents with Cognitive Resource Self-Allocation

What it claims

CORAL gives a ReAct-style web agent two extra tools. One writes a persistent plan-and-facts memory. The other wipes the conversation history and restarts from that memory alone. Training uses SFT followed by Multi-episode ARPO, an extension of DAPO. The paper reports 42.8% average accuracy on GAIA with an 8B model and 6-7x fewer input tokens than a ReAct agent. The authors say most of the GAIA gains come on Level 2 and Level 3, the longest tasks. The paper is accepted at COLM 2026 (arXiv v1, 29 October 2026). Authors are from CUHK, Tencent and HKUST.

How it works

• Memory Management (MM) keeps two lists: task progress (the plan, with each sub-goal marked pending or completed) and verified facts (intermediate results, discovered entities). The agent can add, modify or delete entries. The authors observe three typical calls: at task start to record a plan, on finding critical information, and after finishing a sub-task.
• Context Optimization (CO) does a hard reset. The next episode's context is only the system prompt, the question and the current memory. All tool outputs and reasoning traces are discarded.
• Who decides. The agent calls both tools on its own judgment. No fixed interval or context-saturation trigger is used, which is the stated difference from prior pruning and compression methods.

flowchart LR
  A[System prompt + task + memory] --> B[Thought / action / observation]
  B --> C{Agent's choice}
  C -->|search / browse| B
  C -->|MM: update memory| B
  C -->|CO: reset| A
  C -->|answer| D[Final answer]

SFT stage. Each trajectory is split into its episodes, and each episode is a separate training instance with its own initial context. Loss is masked on observation tokens. Non-terminal episodes teach when to call CO. Later episodes teach resuming from a checkpoint without redoing completed work. The framework also runs prompting-only.

Multi-episode ARPO. For each question, G trajectories are sampled. The reward comes from the final episode and is copied to every earlier episode of that trajectory. Advantages are normalized across all episodes for the question, with DAPO's asymmetric clipping. The reward is a format score (valid trajectory format and valid JSON tool calls) multiplied by an LLM-judge answer score.

Evidence

• Setup. The agent has Web Search and Web Browse plus MM and CO. Three variants are tested: Prompting, SFT, and SFT+RL.
• GAIA. Evaluation uses the 103-question text-only validation subset (39 Level 1, 52 Level 2, 12 Level 3).
• Other benchmarks. Knowledge QA covers NQ, HotpotQA, 2WikiMultiHopQA, Musique and Bamboogle. BrowseComp is also tested; there, CORAL added to an unnamed proprietary agent gains 8.0 points.
• Baselines. These include GPT-4o, DeepSeek-R1 and Claude Sonnet 4 used directly. Trained agents include R1-Searcher, WebDancer, WebThinker, Cognitive Kernel-Pro, WebSailor, WebShaper and OpenAI Deep Research. The authors note these agents are themselves trained (multi-stage RL, SFT with DPO, or SFT with RL).
• Training data. Questions are built with the WebShaper formalization-driven method, at controllable difficulty.

Limits

The fetched text stops in the dataset section, so the per-variant tables, per-level GAIA scores, ablations and Knowledge QA numbers are not visible. The text does not state the 8B base model, the training data size, the compute used, or the judge model. It gives no code, weights or licence. The GAIA figure rests on 103 questions, only 12 of them Level 3.