arxiv arXiv cs.LG · 1h ago · research

ReFold training-free reversible context folding for long-horizon agents

from English

Researchers introduce ReFold, a training-free rendering layer that preserves interaction history while compressing the model's rendered context for long-horizon LLM agents. It removes inter-turn redundancy by replacing displayed content with stubs and folding finished turns into notes, using chunked rewriting to maintain prefix caches.

Removes content already shown in earlier turns and folds agent-reported finished turns into one-line notes.
Uses chunked rendering to rewrite the cached prefix once every few steps instead of at every step.
Ensures all removals are strictly reversible, allowing restoration from history if a removal is incorrect.
Reduces token consumption by up to 2.5x and halves KV-cache memory per session without degrading task success rates.
Avoids up to 92% of forced compactions under capped context budgets and reduces request queuing delays by up to 100%.

ReFold operates as a plug-and-play solution across standard ReAct-style harnesses, accelerating inference by up to 1.7x while cutting inference costs by up to 3.4x.

Importance 2/3 New harness with differentiators arXiv cs.LG AI agents Inference efficiency

Related articles

arxiv arXiv cs.CL · 3h ago · 1 view

ReFold training-free reversible context folding for long-horizon agents

Researchers introduce ReFold, a training-free rendering layer that compresses the model's rendered context in long-horizon LLM agents without discarding the underlying interaction history. It removes inter-turn redundancy by replacing displayed content with stubs and folding finished turns into notes, using chunked rendering to rewrite the cached prefix periodically.

arxiv arXiv cs.CL · 12d ago · 29 views

Coroutine-bridge harness wins CAR-bench Track 2 with 60.0% Pass^3

A new coroutine-bridge harness for tool-using agents won Track 2 of the CAR-bench evaluation, achieving a 60.0% Pass@3 score on the official hidden test set. The system decouples model invocation from tool round-trips by having the model emit Python programs that block and resume across evaluator exchanges.

arxiv arXiv cs.AI · 14d ago · 48 views

CliffCompaction reduces coding agent costs by 50% while maintaining performance

Researchers introduce CliffCompaction, an autocompaction technique designed to reduce costs for long-horizon coding agents by up to 50% under a bounded context. The method maintains or improves performance on Terminal-Bench and achieves state-of-the-art results on KernelBench by keeping compacted information faithful through truncation rather than rephrasing.

arxiv arXiv cs.LG · 14d ago · 44 views

CliffCompaction reduces long-horizon coding agent costs by up to 50% while maintaining performance

Researchers introduce CliffCompaction, an autocompaction technique designed for agents handling millions of tokens, which cuts costs by up to 50% under bounded context. The method maintains or improves performance on Terminal-Bench and achieves state-of-the-art results on KernelBench by keeping information faithful through truncation rather than rephrasing.

lab OpenAI News · 14d ago · 55 views

Parallel halves research time and cost using GPT-6 Astra

Parallel has integrated OpenAI's GPT-6 Astra into its AI agent infrastructure, achieving a 50% reduction in both research time and code costs compared to prior models. The company reports that the new model allows agents to complete complex knowledge work significantly faster while maintaining high-quality output.