ThinkReset compresses reasoning state to extend LLM long‑chain performance
The paper introduces ThinkReset, a learnable intermediate interface that writes back a compressed reasoning snapshot and resets the model’s context window. By optimizing for post‑reset continuation, it cuts redundancy and prevents error buildup, boosting success rates on long‑horizon benchmarks while staying within fixed context limits.