Search papers, labs, and topics across Lattice.
The paper introduces PACE, a dual-level framework for compressing reasoning traces in Language Reasoning Models (LRMs) by addressing overthinking and excessive token usage. PACE employs prefix-protected optimization at the sequence level using decaying mixed rollouts to preserve valid reasoning paths while encouraging conciseness, and difficulty-aware penalty at the group level to dynamically adjust length constraints based on query complexity. Experiments on DeepSeek-R1-Distill-Qwen models (1.5B/7B) demonstrate that PACE achieves up to 55.7% token reduction and up to 4.1% accuracy improvement on math benchmarks, generalizing to code, science, and general domains.
Achieve up to 55.7% token reduction and 4.1% accuracy improvement in language reasoning by selectively compressing reasoning traces, proving that less can be more.
Language Reasoning Models (LRMs) achieve strong performance by scaling test-time computation but often suffer from ``overthinking'', producing excessively long reasoning traces that increase latency and memory usage. Existing LRMs typically enforce conciseness with uniform length penalties, which over-compress crucial early deduction steps at the sequence level and indiscriminately penalize all queries at the group level. To solve these limitations, we propose \textbf{\model}, a dual-level framework for prefix-protected and difficulty-aware compression under hierarchical supervision. At the sequence level, prefix-protected optimization employs decaying mixed rollouts to maintain valid reasoning paths while promoting conciseness. At the group level, difficulty-aware penalty dynamically scales length constraints based on query complexity, maintaining exploration for harder questions while curbing redundancy on easier ones. Extensive experiments on DeepSeek-R1-Distill-Qwen (1.5B/7B) demonstrate that \model achieves a substantial reduction in token usage (up to \textbf{55.7\%}) while simultaneously improving accuracy (up to \textbf{4.1\%}) on math benchmarks, with generalization ability to code, science, and general domains.