@HuggingPapers: Cut your losses in parallel reasoning STOP learns to prune doomed trajectories early by reading KV-cache states, cuttin…
Summary
STOP method prunes doomed reasoning trajectories early via KV-cache states, cutting token usage 70% and boosting AIME/GPQA accuracy across 1.5B–20B models.
View Cached Full Text
Cached at: 04/21/26, 10:33 AM
Cut your losses in parallel reasoning STOP learns to prune doomed trajectories early by reading KV-cache states, cutting token usage by 70% while boosting accuracy on AIME and GPQA across models from 1.5B to 20B parameters.
Similar Articles
Cut Your Losses! Learning to Prune Paths Early for Efficient Parallel Reasoning
This paper introduces STOP (Super Token for Pruning), a lightweight method that learns to prune unpromising reasoning paths early during parallel decoding by appending learnable tokens and reading KV cache states, achieving 70% token reduction while improving performance on AIME and GPQA benchmarks.
Cut Your Losses! Learning to Prune Paths Early for Efficient Parallel Reasoning
This paper proposes STOP (SuperTOken for Pruning), a systematic framework for pruning inefficient reasoning paths early in parallel reasoning with Large Reasoning Models. The method achieves superior efficiency and effectiveness across models from 1.5B to 20B parameters, boosting GPT-OSS-20B accuracy on AIME25 from 84% to 90% under fixed compute budgets.
When Does Learning to Stop Help? A Cost-Aware Study of Early Exits in Reasoning Models
This paper introduces LearnStop, a lightweight checkpoint stopper for reasoning models that predicts prefix correctness from online features, and finds that learned stopping provides value over scalar rules only when many questions become correct early without a single reliable scalar signal.
Stop When Reasoning Converges: Semantic-Preserving Early Exit for Reasoning Models
This paper introduces PUMA, a plug-and-play framework that detects semantic redundancy in chain-of-thought reasoning to enable early exit, achieving 26.2% average token reduction across multiple models and benchmarks while preserving accuracy and reasoning quality.
Thinking Past the Answer: Evaluating Harmful Overthinking in Large Reasoning Models
This paper introduces a prefix-level trajectory evaluation protocol to distinguish harmful overthinking from verbose but harmless overthinking in large reasoning models, showing that continued reasoning after reaching the correct answer can destabilize performance. The authors find that early stopping improves accuracy by up to 21% on multimodal benchmarks, and identify logical drift and visual reinterpretation as key causes of correctness deviations.