Think Right: Learning to Mitigate Under-Over Thinking via Adaptive, Attentive Compression

• URL: http://arxiv.org/abs/2510.01581v1
• Date: Thu, 02 Oct 2025 02:00:20 GMT
• Title: Think Right: Learning to Mitigate Under-Over Thinking via Adaptive, Attentive Compression
• Authors: Joykirat Singh, Justin Chih-Yao Chen, Archiki Prasad, Elias Stengel-Eskin, Akshay Nambi, Mohit Bansal,
• Abstract summary: We propose an online post-training RL method that prunes redundant steps and estimates difficulty.<n> TRAAC (Think Right with Adaptive, Attentive Compression) achieves an average absolute accuracy gain of 8.4%.<n>Although our models are trained on math datasets, they show accuracy and efficiency gains on out-of-distribution non-math datasets.
• Score: 68.69801176669843
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Recent thinking models solve complex reasoning tasks by scaling test-time compute, but this scaling must be allocated in line with task difficulty. On one hand, short reasoning (underthinking) leads to errors on harder problems that require extended reasoning steps; but, excessively long reasoning (overthinking) can be token-inefficient, generating unnecessary steps even after reaching a correct intermediate solution. We refer to this as under-adaptivity, where the model fails to modulate its response length appropriately given problems of varying difficulty. To address under-adaptivity and strike a balance between under- and overthinking, we propose TRAAC (Think Right with Adaptive, Attentive Compression), an online post-training RL method that leverages the model's self-attention over a long reasoning trajectory to identify important steps and prune redundant ones. TRAAC also estimates difficulty and incorporates it into training rewards, thereby learning to allocate reasoning budget commensurate with example difficulty. Our approach improves accuracy, reduces reasoning steps, and enables adaptive thinking compared to base models and other RL baselines. Across a variety of tasks (AIME, AMC, GPQA-D, BBEH), TRAAC (Qwen3-4B) achieves an average absolute accuracy gain of 8.4% with a relative reduction in reasoning length of 36.8% compared to the base model, and a 7.9% accuracy gain paired with a 29.4% length drop compared to the best RL baseline. TRAAC also shows strong generalization: although our models are trained on math datasets, they show accuracy and efficiency gains on out-of-distribution non-math datasets like GPQA-D, BBEH, and OptimalThinkingBench. Our analysis further verifies that TRAAC provides fine-grained adjustments to thinking budget based on difficulty and that a combination of task-difficulty calibration and attention-based compression yields gains across diverse tasks.

Related papers

• Training LLMs for Divide-and-Conquer Reasoning Elevates Test-Time Scalability [129.1296673737603]
  Large language models (LLMs) have demonstrated strong reasoning capabilities through step-by-step chain-of-thought (CoT) reasoning.<n>A potential alternative is divide-and-conquer (DAC) reasoning, which decomposes a complex problem into subproblems to facilitate more effective exploration of the solution.<n>We propose an end-to-end reinforcement learning (RL) framework to enhance their DAC-style reasoning capacity.
  arXiv  Detail & Related papers  (2026-02-02T18:54:54Z)
• TACLer: Tailored Curriculum Reinforcement Learning for Efficient Reasoning [24.04507196574173]
  Large Language Models (LLMs) have shown remarkable performance on complex reasoning tasks.<n>We propose TACLer, a model-tailored curriculum reinforcement learning framework.<n> TACLer features two core components: (i) tailored curriculum learning that determines what knowledge the model lacks and needs to learn in progressive stages; (ii) a hybrid Thinking/NoThinking reasoning paradigm that balances accuracy and efficiency by enabling or disabling the Thinking mode.
  arXiv  Detail & Related papers  (2026-01-29T13:40:35Z)
• Correct, Concise and Complete: Multi-stage Training For Adaptive Reasoning [11.179446105672461]
  We propose a multi-stage efficient reasoning method that combines supervised fine-tuning and reinforcement learning.<n>Our approach reduces response length by an average of 28% for 8B models and 40% for 32B models.<n>It achieves a superior trade-off compared to more complex state-of-the-art efficient reasoning methods.
  arXiv  Detail & Related papers  (2026-01-06T12:31:51Z)
• DART: Difficulty-Adaptive Reasoning Truncation for Efficient Large Language Models [36.962276192354174]
  textbfDART adjusts thinking length according to problem difficulty.<n>textbfTruncation framework learns when to stop thinking''
  arXiv  Detail & Related papers  (2025-11-03T02:41:20Z)
• e1: Learning Adaptive Control of Reasoning Effort [88.51897900019485]
  Increasing the thinking budget of AI models can significantly improve accuracy, but not all questions warrant the same amount of reasoning.<n>Users may prefer to allocate different amounts of reasoning effort depending on how they value output quality versus latency and cost.<n>We propose Adaptive Effort Control, a self-adaptive reinforcement learning method that trains models to use a user-specified fraction of tokens.
  arXiv  Detail & Related papers  (2025-10-30T23:12:21Z)
• DLER: Doing Length pEnalty Right - Incentivizing More Intelligence per Token via Reinforcement Learning [134.03095505580276]
  Doing Length pEnalty Right (DLER) is a training recipe combining batch-wise reward normalization, higher clipping, dynamic sampling, and a simple truncation length penalty.<n>DLER achieves state-of-the-art accuracy--efficiency trade-offs, cutting output length by over 70 percent while surpassing all previous baseline accuracy.
  arXiv  Detail & Related papers  (2025-10-16T20:05:57Z)
• Staying in the Sweet Spot: Responsive Reasoning Evolution via Capability-Adaptive Hint Scaffolding [59.60915947702282]
  Reinforcement learning with verifiable rewards (RLVR) has achieved remarkable success in enhancing the reasoning capabilities of large language models (LLMs)<n>Existing RLVR methods often suffer from exploration inefficiency due to mismatches between the training data's difficulty and the model's capability.<n>We propose SEELE, a novel supervision-aided RLVR framework that dynamically adjusts problem difficulty to stay within the high-efficiency region.
  arXiv  Detail & Related papers  (2025-09-08T17:36:21Z)
• Less is More Tokens: Efficient Math Reasoning via Difficulty-Aware Chain-of-Thought Distillation [82.2288581878096]
  We present a framework for difficulty-aware reasoning that teaches models to dynamically adjust reasoning depth based on problem complexity.<n>We show that models can be endowed with such dynamic inference pathways without any architectural modifications.
  arXiv  Detail & Related papers  (2025-09-05T16:40:13Z)
• CoThink: Token-Efficient Reasoning via Instruct Models Guiding Reasoning Models [56.40065909544213]
  Large language models (LLMs) benefit from increased test-time compute, a phenomenon known as test-time scaling.<n>However, reasoning-optimized models often overthink even simple problems, producing excessively verbose outputs and leading to low token efficiency.<n>We identify two key causes of this verbosity: (1) reinforcement learning reduces the information density of forward reasoning, and (2) backward chain-of thought training encourages redundant and often unnecessary verification steps.
  arXiv  Detail & Related papers  (2025-05-28T06:24:45Z)
• LIMOPro: Reasoning Refinement for Efficient and Effective Test-time Scaling [29.721108461390973]
  We introduce PIR (Perplexity-based Importance Refinement), a principled framework that quantitatively evaluates the importance of each reasoning step.<n>PIR identifies and selectively prunes only low-importance functional steps while preserving progressive reasoning components.<n>Our approach demonstrates strong generalizability across different model sizes, data sources, and token budgets.
  arXiv  Detail & Related papers  (2025-05-25T15:17:57Z)
• ConciseRL: Conciseness-Guided Reinforcement Learning for Efficient Reasoning Models [14.403953640255823]
  We introduce a novel score used as a reward signal within a reinforcement learning framework to guide models toward generating correct and concise reasoning traces.<n>This score is evaluated by a large language model acting as a judge, enabling dynamic, context-aware feedback beyond simple token length.<n>Our method achieves state-of-the-art efficiency-accuracy trade-offs on the MATH dataset, reducing token usage by up to 31x on simple problems while improving accuracy by 7%, and on the hardest problems, it outperforms full reasoning by +7.5% accuracy with up to 3.6x fewer tokens.
  arXiv  Detail & Related papers  (2025-05-22T19:56:35Z)
• Plan and Budget: Effective and Efficient Test-Time Scaling on Large Language Model Reasoning [19.258292534503887]
  Plan-and-Budget is a model-agnostic, test-time framework that decomposes complex queries into sub-questions and allocates token budgets based on estimated complexity using adaptive scheduling.<n>Plan-and-Budget improves reasoning efficiency across a range of tasks and models, achieving up to +70% accuracy gains, tangential -39% token reduction, and +187.5% improvement in $E3$.
  arXiv  Detail & Related papers  (2025-05-22T01:56:29Z)
• THOUGHTTERMINATOR: Benchmarking, Calibrating, and Mitigating Overthinking in Reasoning Models [65.39456695678713]
  We introduce approximate measures of problem-level difficulty and demonstrate that a clear relationship between problem difficulty and optimal token spend exists.<n>We find that in general, reasoning models are poorly calibrated, particularly on easy problems.<n>We introduce THOUGHTTERMINATOR, a training-free black box decoding technique that significantly improves reasoning model calibration.
  arXiv  Detail & Related papers  (2025-04-17T22:16:30Z)
• DCR: Divide-and-Conquer Reasoning for Multi-choice Question Answering with LLMs [9.561022942046279]
  We propose Divide and Conquer Reasoning (DCR) to enhance the reasoning capability of large language models (LLMs) We first categorize questions into two subsets based on confidence score ($mathcalCS$), which is estimated by statistical frequency of generated answers. In particular, we first categorize questions into two subsets based on confidence score ($mathcalCS$), which is estimated by statistical frequency of generated answers.
  arXiv  Detail & Related papers  (2024-01-10T14:38:46Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.