ThinkReset: Learnable Intermediate Interface Construction for Bounded-Context Long-Horizon Reasoning
06:00 · August 3, 2026 · arXiv cs.AI RSS
Long chain-of-thought reasoning improves performance on complex problems, but it also introduces redundancy accumulation, context overflow, and error anchoring. We argue that under bounded context windows, the core bottleneck is not trajectory compression or test-time control, but the absence of a reusable intermediate interface that can replace discarded history and support continued solving. We further identify a key failure mode of outcome-reward-driven long-chain reinforcement learning: when the model has not solved the task before the window is nearly exhausted, the final-answer reward encourages premature guessing rather than continued careful reasoning. We propose ThinkReset, a text-space instantiation of this view. ThinkReset explicitly constructs reusable intermediate interfaces through interface writeback and reset, and directly optimizes post-reset continuation success. Across multiple long-horizon reasoning benchmarks, this perspective consistently improves success rates under fixed context windows.
Summary
ThinkReset addresses a core limitation in long chain-of-thought reasoning for large language models: as traces lengthen under fixed context windows, redundancy accumulates, incorrect hypotheses persist, and unfinished solution processes risk truncation. Rather than treating these issues primarily as problems of trajectory compression or test-time scheduling, the approach reframes bounded-context long-horizon reasoning as an intermediate interface learning task. The central requirement is to produce a reusable textual state that can replace discarded history while still enabling continued problem solving.
The method implements this view through explicit interface writeback and reset. Once context usage reaches a threshold, the model generates a compact intermediate representation and substitutes it for the preceding trace. Training proceeds in three stages: an initial cold-start supervised fine-tuning phase, followed by reinforcement learning with RLOO that rewards successful continuation after reset, and a final reset-specific training pass that further aligns the interface with post-reset solvability. This objective directly counters a documented failure mode in outcome-reward-driven long-chain reinforcement learning, where models shift toward premature guessing as the window fills rather than preserving reasoning capacity.
Experiments apply the framework to Qwen3 models on AIME, ZebraLogic, AutoLogi, and GPQA. Under a fixed 32k context window the method yields consistent gains in success rate and reasoning stability compared with trajectory-retention baselines, confirming that optimizing for reusable intermediate interfaces improves long-horizon performance without requiring additional context capacity.
Why it matters
Provides actionable, reproducible techniques for efficient long-horizon reasoning that Dutch AI researchers and advanced practitioners can implement to improve LLM performance within EU context-length and compute constraints.
