ThinkReset Improves LLM Long-Horizon Reasoning with Context Management

Long chain-of-thought reasoning, while effective for complex problems, often suffers from issues like redundancy, context window overflow, and the anchoring of errors. This research posits that the primary bottleneck in bounded-context scenarios is the lack of a reusable intermediate interface that can effectively replace discarded historical context and enable continued problem-solving. The study also identifies a critical failure mode in outcome-reward-driven reinforcement learning for long chains: when the context window is nearly exhausted before a solution is found, the model is incentivized to guess prematurely rather than continue careful reasoning. To address these challenges, the researchers propose ThinkReset. ThinkReset explicitly constructs these reusable intermediate interfaces through a process of interface writeback and reset. Crucially, it directly optimizes for the success of the reasoning process *after* a reset, ensuring that the model can pick up from a well-summarized state. This approach consistently led to improved success rates across various long-horizon reasoning benchmarks, even when operating within fixed context windows.