Large Reasoning Models (LRMs) have demonstrated remarkable capabilities by scaling up the length of Chain-of-Thought (CoT). However, excessively long reasoning traces pose substantial challenges for training cost and inference latency. While various CoT compression approaches have emerged to address this challenge, they face inherent trade-offs: token-level methods often disrupt syntactic and logical coherence, while step-level methods based on perplexity fail to reliably capture the logically critical reasoning steps because of the dilution of logical information. In this paper, we propose ASAP (Anchor-guided, SurprisAl-based Pruning), a novel coarse-to-fine framework for CoT compression. ASAP first performs anchor-guided pruning to preserve the core reasoning structure, which efficiently reduces the search space for subsequent processing. Leveraging the insight that logical branching choices are concentrated at the onset of reasoning steps, it then enables logic-aware pruning by selecting logically essential reasoning steps based on a novel first-token surprisal metric. Finally, ASAP distills these concise CoTs into target models so they can generate efficient reasoning traces at inference time. Experiments across math and code benchmarks show that ASAP preserves competitive accuracy while substantially reducing generated tokens, latency, and training cost, with additional validation on Qwen3 confirming its cross-model applicability.

Interactive large language model agents are often specialized to a single environment and transfer poorly across others. We study training-free consolidation of such specialists into a single generalist checkpoint. A key difficulty in this setting is that, in multi-turn agents, small merge errors on role-critical outputs can trigger long-horizon failure cascades. We therefore propose Agent-Role Merging (ARM), a neuron-level training-free framework that selects a merged backbone with an answer-free activation-overlap signal and then repairs residual gaps through conflict-aware transplantation of a small set of role-salient neurons. Our analyses further show that role-conditioned tracing yields more benchmark-specific neuron sets, enabling more localized repair. Across Qwen3-8B and Qwen2.5-7B expert pools, ARM produces the strongest balanced merged generalist among compared training-free baselines, improving both average performance and worst-suite robustness while preserving out-of-domain generalization. These results suggest that targeted neuron-level edits offer a practical route to robust multi-turn agent merging without end-to-end retraining.

The rapid growth of LLM evaluation resources makes exhaustive benchmark curation increasingly difficult. Existing efficient benchmarking methods reduce evaluation cost by selecting representative subsets or predicting full-benchmark performance, but typically require evaluation results from many models and additional training. In this paper, we ask a more fundamental question: to what extent do benchmarks actually measure different things? We study this question at two levels: intra-benchmark redundancy, which supports benchmark compression, and inter-benchmark similarity, which supports evaluation-suite design. We first formalize both notions through a functionalist behavioral criterion, and then propose an evaluation-free proxy based on Sparse Autoencoder (SAE) features for redundancy diagnosis. Our central hypothesis is that behavioral overlap can be approximated by feature overlap. This hypothesis yields level-specific metrics based on feature-coverage saturation within benchmarks and feature-footprint overlap across benchmarks. Experiments show that SAE feature-based redundancy strongly correlates with performance-based redundancy across 17 benchmarks, while SAE-guided subset selection produces more stable and representative subsets than random sampling. SAE feature overlap also predicts inter-benchmark performance similarity across 8 benchmarks. These results show that SAE features can turn benchmark analysis from a costly performance-matrix problem into a feature-footprint problem, enabling evaluation-free benchmark compression and suite curation.