May 2026· arXiv.org· Vol abs/2605.17026· 1 citation· 43 references
Computer Science
TL;DR
This paper investigates the cause of reasoning shrinkage under SFT-based post-training and identifies data-centric factors as a key driver of shrinkage in reasoning models and highlights diversity-aware designs as an effective lever for controlling it.
Abstract
Recent progress in large language models has led to the emergence of reasoning models, which have shown strong performance on complex tasks through specialized fine-tuning procedures. While these methods reliably improve pass@1 accuracy, prior works have observed that they show a coverage shrinkage behavior, where pass@k degrades relative to the base model. In this paper, we investigate the cause of reasoning shrinkage under SFT-based post-training. We hypothesize that this behavior is driven by properties of the fine-tuning data, specifically related to decision points or"forks in the road"scenarios where model encounters indecipherable patterns with multiple valid reasoning paths. To test this hypothesis, we design controlled case studies that simulate such decision-point settings, spanning indecipherable nodes in graph branching, and reasoning modes. By tracking post-training dynamics in these settings, we find that the shrinkage phenomenon is tightly correlated with the prevalence of decision-point scenarios in the training data. We also demonstrate that this shrinkage behavior can be partially mitigated through targeted data synthesis design of decision-points and a more systematic diversity-encouraging decoding mechanism. Our findings identify data-centric factors as a key driver of shrinkage in reasoning models and highlight diversity-aware designs as an effective lever for controlling it. (Data and code for reproducing our experiments are available at https://github.com/psunlpgroup/reasoning_forks )
It is found that full trajectories provide only limited benefit, while partial trajectories are effective even under heavy truncation, and training LLMs using endpoints leads to consistent changes in reasoning behavior, and that it also benefits post-training methods based on reinforcement learning or on-policy distill...
Jaehui Hwang, Sangdoo Yun, Byeongho Heo et al.· 0 citations
Large reasoning models (LRMs) generate long chain-of-thought traces before answering, creating significant inference overhead. Pruning can reduce this cost, but its effectiveness depends on the calibration data used to estimate parameter importance. Recent work calibrates on the model's own rollouts instead of generic...
Ha Lan Nguyen, Huy Hoang Son Tran, Trac-Duy Tran et al.· 1 citation
A deformation-controlled, support-preserving Power target is replaced with a deformation-controlled, support-preserving Power target that calibrates sharpening across problems while limiting the suppression of moderate-probability paths to solve a striking paradox.
The Powered Length Penalty (PLP) is proposed, an adaptive regularizer that penalizes redundancy in short sequences while gradually reducing penalties for longer sequences, preserving deep reasoning.
Ze-Hui Ling, De-Shu Chen, Hong-Wei Zhang et al.· Proceedings of the Thirty-Fi...· 0 citations
STRAT is introduced, which partitions the residual stream into orthogonal Data and Type subspaces and uses Type-based attention and gating to govern Data transformations and in arithmetic, STRAT reduces median OOD error 35-fold relative to a Transformer baseline.
Cristina V. Lopes, Yuan-Gang Li, Justin Tian Jin Chen et al.· arXiv.org· 0 citations
The Distributional Answer Commitment Score (DACS) is proposed, a distributional confidence estimator that measures the entropy of the model's probability distribution over answer commitment at each reasoning step to capture how concentrated the model's answer belief is without requiring ground-truth answers or task-spe...
Zhao-Han Zhang, Jun-Jie Liu, Chengzhengxu Li et al.· arXiv.org· 0 citations
Related blog posts
MIT News · Artificial Intelligence· news.mit.eduNov 17, 2026
By studying the computational foundations of language learning, Associate Professor Jacob Andreas builds intelligent systems that can communicate effectively with humans.
Microsoft Research Blog· microsoft.comNov 16, 2026
AI coding agents are increasingly help developers fix bugs and build features largely on their own, but every interaction with their models consumes tokens. In each session, an agent sends instructions, tool descriptions, and session history to its model. Most efforts to reduce this cost focus on individual tasks. Yet developers return to the same […] The post TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient appeared first on Microsoft Research.
MIT News · Artificial Intelligence· news.mit.eduNov 13, 2026