What is CORAL in LLM agents?

CORAL (Cognitive Resource Self-Allocation) is a framework that gives an agent two tools: **Memory Management** to checkpoint its plan and verified facts, and **Context Optimization** to clear its conversation history and restart from only that checkpoint. The agent decides when to call each one.

Why does resetting the context cut input tokens?

A ReAct loop re-reads its whole history on every turn, so input cost grows quickly with turn count. In the paper's GAIA runs with Qwen3-8B, resets capped peak context at 12.7K tokens against 30.6K, and the saved plan cut LLM calls from 45.3 to 15.2 per trajectory.

How is CORAL different from context compaction?

Compaction rewrites or summarizes the history, usually when a rule or an outside process triggers it. CORAL lets the agent itself choose when to reset, and it carries forward only an explicit checkpoint of task progress and verified facts, not a summary of the conversation.

What are the risks of an agent-controlled context reset?

A reset discards the history that could help the agent repair an earlier mistake, so a fact wrongly saved as verified can persist. The paper also finds that prompting-only CORAL helps strong backbones but hurt Qwen2.5-7B-Instruct by 3.9 points, because it could not reliably call four tools.

Agent· 2026-10-31

Checkpoint agent context to use 6–7× fewer input tokens — Agent-callable context reset — What does it mean?

Jargon

ReAct: The standard agent loop: the model alternates a reasoning step, a tool call and the tool's result, and every step stays in the context until the task ends. See the agent loop.
Episode: One stretch of reasoning between two context resets (or between the last reset and the end of the task). A CORAL task is a chain of episodes, and each later episode starts from the system prompt, the question and the saved memory only.
Checkpoint: The agent's saved working memory, in two parts: task progress (the plan and which sub-goals are done) and verified facts (results and entities it has confirmed). It survives a reset.
Context reset (CO): The Context Optimization tool. It wipes the conversation history and returns the checkpoint as the first observation of a fresh episode.
Multi-episode ARPO: The paper's reinforcement-learning method. It extends the DAPO policy-gradient method so that each episode is its own training unit, while all episodes of one trajectory share the reward computed from the final answer.
GAIA: A benchmark of real-world questions that need web browsing, tool use and multi-step reasoning. Level 1 is easiest and Level 3 is the longest-horizon.
BrowseComp: A much harder web-research benchmark whose questions need many hops of exploration before an answer can be found.

The news. On October 29, 2026, researchers from CUHK, HKUST and Tencent posted "Don't Lose the Thread", which introduces CORAL. With an 8B model (Qwen3-8B) it reaches 42.8% average accuracy on GAIA, and its abstract reports 6–7× fewer input tokens than a ReAct agent. With no training at all, it also lifts Claude Sonnet 4 by 8.0 points on BrowseComp. Read the paper →

Picture a researcher working a hard question at a desk. Every page she pulls up stays on the desk, and after thirty pages the one fact she needs is under the pile. A ReAct agent's context window is that desk: every tool output and old thought stays on it, and the model re-reads the whole pile before each new step. The paper calls the result cognitive overload: the irrelevant history dilutes attention and planning gets worse.

The usual fixes are done to the agent: a rule fires at a token limit, or an outside process compacts the history. CORAL's move is different. It hands the agent two tools and lets the agent decide, from its own read of the task state, when to write things down and when to clear the desk. The researcher writes a plan and confirmed facts on an index card, sweeps the desk clean, and starts again with only the card.

The two tools are small. Memory Management (MM) adds, edits or deletes entries in two sections: task progress (the plan, and which sub-goals are finished) and verified facts (confirmed intermediate results or entities). The agent typically calls it at three moments: to record a plan at the start, to store a critical discovery, and after finishing a sub-task. Context Optimization (CO) is the hard reset. The history is cleared, the memory is returned as an observation, and the next episode begins from (system prompt, question, memory), so only checkpointed information survives; accumulated tool outputs and reasoning traces are discarded.

Workreason, search, call MM

Judge clutteragent decides

Call COhistory cleared

Resumeprompt + question + card

The framework works with prompting alone, but the paper also trains for it. Because a trajectory is now several episodes with different starting contexts, it is split into one training example per episode, and the loss skips observation tokens so the model learns only its own thoughts and calls. Early episodes teach when a batch of sub-tasks is done and the context should be discarded; later episodes teach resuming from a compact card instead of a history. For reinforcement learning, the multi-episode ARPO method computes the reward from the final episode and shares it with every earlier episode, which gives credit assignment across a reset. This links to the cost side of context as a scarce resource.

The cost saving has two separate factors, and the paper's Table 2 lets you see both. On GAIA with Qwen3-8B, a ReAct trajectory averages 45.3 LLM calls and 1,078K input tokens, which is about 23.8K tokens re-read per call (derived by dividing the two averages). The RL-trained CORAL agent averages 15.2 calls and 136K input tokens, about 8.9K per call. The agent makes about 3× fewer calls because it stops re-searching facts it already wrote down, and each call carries about 2.7× less context because resets cap the window (peak 12.7K tokens versus 30.6K). Multiplied, that is about 7.9× fewer input tokens, or 87.4% fewer (rounded from the table averages). The abstract states the headline as 6–7×; the Table 2 averages for the RL model work out slightly higher.

Method (Qwen3-8B, GAIA)	Input tokens per trajectory	LLM calls	Peak context	GAIA avg. accuracy
ReAct, Table 2	1,078K	45.3	30.6K	19.4%
CORAL, prompting only, Table 2	314K	21.8	14.3K	20.4%
CORAL, SFT, Table 2	153K	15.9	13.9K	41.6%
CORAL, RL, Table 2	136K	15.2	12.7K	42.8%

Read the table for what it says about training. Prompting alone cuts tokens by about 3.4× but barely moves accuracy (20.4% against 19.4%); the accuracy jump to 41.6% and 42.8% comes from SFT and RL. The trained agent averages only 0.21 CO calls per trajectory, against 4.3 for the prompted one. Training did not teach it to reset more; it taught it to checkpoint better, so the clutter that would justify a reset never builds up.

Several caveats are in the paper itself. A reset also throws away the history an agent could use to repair an earlier mistake, so a wrong fact saved as "verified" can carry into later episodes. The prompting-only benefit depends on the backbone: it helps stronger models such as QwQ-32B (+5.8) and Claude Sonnet 4 (+4.8), is neutral on Qwen3-8B, and costs Qwen2.5-7B-Instruct 3.9 points because it could not reliably emit calls for four tools instead of two. The BrowseComp result (30.7% against 22.7%) covers only the first 150 BrowseComp questions, the subset the authors could afford to run, with Claude Sonnet 4 and identical search tools. Compared with compaction, which rewrites history, checkpointing is a different bet: save a few verified facts and drop the rest of the history entirely.

Goes deeper in: AI Agents → Context Engineering → The 4 Fixes

Frequently Asked Questions

Check what you knowMap your AI & GPU knowledge across every track — free, role-based