Multi-objective Evolutionary Merging Enables Efficient Reasoning Models

Large reasoning models (LRMs) achieve strong performance via extended chain-of-thought (CoT) reasoning, yet suffer from excessive token consumption and high inference latency. Existing reinforcement learning (RL) approaches for CoT compression rely on uniform, static length penalties that neglect model capability dynamics and problem-level difficulty variation. We propose ExpThink\xspace, an RL framework that addresses both dimensions through two complementary mechanisms. First, experience-guided reward shaping tracks the shortest correct solution found so far for each problem and applies a three-tier reward: full credit for concise correct responses, discounted credit for verbose correct ones, and zero for incorrect ones. The threshold tightens automatically with model improvement, forming a self-evolving curriculum that requires no manual scheduling. Second, difficulty-adaptive advantage replaces standard deviation normalization with correct-count normalization, yielding monotonically difficulty-scaled gradients that amplify learning on hard problems to preserve accuracy while suppressing gradients on easy ones to encourage brevity. Together, these mechanisms enforce an accuracy-first, compression-second training objective. Experiments on multiple mathematical reasoning benchmarks demonstrate that ExpThink\xspace reduces average response length by up to 77% while simultaneously improving accuracy, achieving up to $3\times$ higher accuracy-efficiency ratio (accuracy divided by average token count) than the vanilla baseline and outperforming existing RL-based compression methods on both metrics.

Reasoning in Large Language Models incurs significant inference-time compute, yet the token-level information structure of reasoning traces remains underexplored. We observe that reasoning tokens split into two functional types: low-entropy structural tokens (recurring phrases that scaffold the reasoning process) and higher-entropy organic tokens (problem-specific content that drives toward a solution). This asymmetry motivates a simple, model-agnostic compression pipeline: apply cross-word BPE merges on a model's own reasoning traces to derive supertokens that capture frequent structural patterns, then teach the model to adopt them via supervised fine-tuning. Across three model families and five mathematical reasoning benchmarks, our approach shortens reasoning traces by 8.1% on average with no statistically significant accuracy loss on any model--benchmark pair. Beyond compression, supertokens act as interpretable reasoning-move annotations (backtracking, verification, strategy shifts), exposing the model's high-level strategy at a glance. Analyzing transitions between structural categories reveals systematic differences between correct and incorrect traces: correct traces show productive recovery (backtracking followed by strategy shifts and verification), while incorrect traces are dominated by confusion cycles (repeated hedging and unresolved contradictions). These diagnostic signals suggest applications in reward shaping and early stopping for RL-based reasoning training.

Large Reasoning Models (LRMs) achieve strong performance on complex tasks by leveraging long Chain-of-Thought (CoT), but often suffer from overthinking, leading to excessive reasoning steps and high inference latency. Existing CoT compression methods struggle to balance accuracy and efficiency, and lack fine-grained, step-level adaptation to redundancy and reasoning bias. Therefore, we propose State-Aware Reasoning Compression with Knowledge Guidance (STACK), a framework that performs step-wise CoT compression by explicitly modeling stage-specific redundancy sources and integrating with a retrieval-augmented guidance. STACK constructs online long-short contrastive samples and dynamically switches between knowledge-guided compression for uncertain or biased reasoning state and self-prompted compression for overly long but confident state, complemented by an answer-convergence-based early stopping mechanism to suppress redundant verification. We further propose a reward-difference-driven training strategy by combining Proximal Policy Optimization (PPO) and Direct Preference Optimization (DPO), enabling models to learn state-conditioned compression strategies. Experiments on three mathematical reasoning benchmarks show that STACK achieves a superior accuracy-efficiency balance, reducing average response length by 59.9% while improving accuracy by 4.8 points over existing methods.