Точка Синхронізації

Researchers recently published a technical paper on the arXiv preprint server introducing 'Free()LM,' a novel framework designed to solve the performance degradation caused by excessive 'thinking' tokens in large language models (LLMs). Released in late February 2024, the study addresses a critical paradox in reasoning models where scaling test-time compute—the amount of processing done during inference—can actually hinder accuracy rather than improve it. The authors argue that current architectures function as 'malloc-only' systems that allocate memory for every computational step but lack the capacity to discard redundant or incorrect logic, leading to cognitive clutter during complex problem-solving tasks. The core of the research identifies a fundamental architectural flaw in how modern AI models handle long-chain reasoning. Unlike traditional software that uses a 'free' command to release unused memory, standard LLMs continuously accumulate both valid and obsolete information within their context window. This accumulation forces the model to sift through an ever-increasing amount of noise, which often results in logical drift or the 'lost in the middle' phenomenon. By failing to prune irrelevant steps, the models eventually collapse under the weight of their own generated tokens, effectively hitting a ceiling of diminishing returns. To overcome these limitations, the proposed Free()LM architecture introduces a mechanism for models to 'learn to forget.' This approach allows the system to actively identify and prune redundant or erroneous intermediate steps during the reasoning process. By mimicking a more human-like cognitive process of filtering out distractions, Free()LM aims to maximize the efficiency of test-time compute. This development represents a significant shift in AI research, moving away from simply increasing the volume of computation toward a more sophisticated management of the model's internal state and memory allocation.