Learning When to Think: Adaptive Reasoning for Test-Time Compute Allocation

• URL: http://arxiv.org/abs/2608.20256v2
• Date: Fri, 21 Aug 2026 07:26:24 GMT
• Title: Learning When to Think: Adaptive Reasoning for Test-Time Compute Allocation
• Abstract summary: We study whether a model can learn to allocate its own reasoning effort by choosing one of three modes.<n>The choice is learned inside Group Relative Policy Optimization with no separate router.<n>We build a reasoning model that adaptively chooses how much to reason for each problem.
• Score: 5.154153910688557
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Reasoning language models trained with reinforcement learning typically operate under a fixed token budget rather than an explicitly adaptive one, which can lead to over-computation on easy problems and insufficient computation on difficult ones. We study whether a model can learn to allocate its own reasoning effort by choosing, as the first token of its response, one of three modes: \textsc{NoThink} (answer as quickly as possible), \textsc{Short} (brief reasoning), or \textsc{Long} (extended reasoning). The choice is learned inside Group Relative Policy Optimization (GRPO) with no separate router, through a shaped reward that makes each mode worthwhile at a different response length, together with hard per-mode token caps that keep the modes distinct. On a 1.5B distilled model trained on MATH, the three modes emerge without collapsing to a single choice, and the brief modes end up more accurate than \textsc{Long}, which shows that the router sorts problems by difficulty rather than at random. Averaged over three seeds, the resulting policy stays close to the base model's accuracy on the held-out MATH500 ($0.782$ vs.\ $0.796$) while cutting the mean response length from $4{,}796$ to $2{,}811$ tokens (a $41\%$ reduction). Interestingly, it also transfers to other benchmarks without retraining, with the largest savings where problems are easier, with for instance 76\% token reduction on GSM8K and at higher accuracy than the baselines at similar response length. In short, we build a reasoning model that adaptively chooses how much to reason for each problem.

Related papers

• Off-the-Shelf LLMs as Process Scorers: Training-Free Alternative to PRMs for Mathematical Reasoning [51.88950852117154]
  Chunk-Level Guided Generation is a training-free alternative that uses an off-the-shelf large language model as a process scorer.<n>We show that scoring variable-length reasoning steps with large-model likelihoods is unreliable due to a systematic length bias.<n>Chunk-Level Guided Generation produces substantially shorter reasoning traces than PRM guided search.
  arXiv  Detail & Related papers  (2026-06-01T04:43:36Z)
• Stateful Reasoning via Insight Replay [51.85629502016196]
  Chain-of-Thought (CoT) reasoning has become a foundation for eliciting multi-step reasoning in large language models.<n>We identify a major cause of this phenomenon: as the CoT grows, the model's attention to critical insights produced earlier in the trace gradually weakens.<n>We propose textbfInsightReplay, a stateful reasoning approach in which the model periodically extracts critical insights from its reasoning trace and replays them near the active generation frontier.
  arXiv  Detail & Related papers  (2026-05-14T06:52:59Z)
• What Do EEG Foundation Models Capture from Human Brain Signals? [64.48249643001402]
  Modern EEG foundation models learn directly from raw signals via self-supervised pretraining.<n>We decompose into three sub-questions: emphwhat does the model learn, emphwhat does the model use, and emphhow much can be explained<n>The audit covers three foundation models (CSBrain, CBraMod, LaBraM), five clinical tasks (MDD, Stress, ISRUC-Sleep, TUSL, Siena), and a 6-family 63-feature lexicon.
  arXiv  Detail & Related papers  (2026-05-12T01:57:53Z)
• The Coupling Tax: How Shared Token Budgets Undermine Visible Chain-of-Thought Under Fixed Output Limits [12.316478346055618]
  Chain-of-thought reasoning is often treated as a monotone way to improve language-model accuracy by letting a model think longer.<n>We identify a countervailing effect, the coupling tax: when reasoning traces and final answers share one output-token budget, long traces can crowd out the answer they are meant to support.<n>Across GSM8K, MATH-500, and five BIG-Bench Hard tasks with Qwen3 models at three scales, non-thinking mode matches or outperforms thinking mode on GSM8K and MATH-500 at every budget up to 2048 tokens.
  arXiv  Detail & Related papers  (2026-05-08T12:54:53Z)
• Shorter but not Worse: Frugal Reasoning via Easy Samples as Length Regularizers in Math RLVR [21.593023093509302]
  We show that retaining and modestly up-weighting moderately easy problems acts as an implicit length regularizer.<n>Experiments using this approach on textitQwen3-4B-Thinking-2507 achieve baseline pass@1 AIME25 accuracy while generating solutions that are, on average, nearly twice as short.
  arXiv  Detail & Related papers  (2025-11-02T17:29:16Z)
• Catch Your Breath: Adaptive Computation for Self-Paced Sequence Production [55.76222360698305]
  We explore a class of supervised training objectives that allow a language model to dynamically and autonomously scale the number of compute steps used for each input token.<n>For any token, the model can request additional compute steps by emitting a don't know> output.<n>We find that the CYB model requests additional steps when doing so improves accuracy, and the model adapts its processing time to token-level complexity and context.
  arXiv  Detail & Related papers  (2025-10-13T21:07:05Z)
• ARM: Adaptive Reasoning Model [52.955386431182724]
  We propose Adaptive Reasoning Model (ARM), a reasoning model capable of adaptively selecting appropriate formats based on the task at hand.<n>Ada-GRPO enables ARM to achieve high token efficiency, reducing tokens by an average of 30%, and up to 70%, while maintaining performance comparable to the model that relies solely on Long CoT.
  arXiv  Detail & Related papers  (2025-05-26T17:38:50Z)
• Concise Reasoning, Big Gains: Pruning Long Reasoning Trace with Difficulty-Aware Prompting [28.537281448659634]
  We propose a difficulty-aware prompting (DAP) method to dynamically shorten reasoning traces without performance loss.<n>In experiments, a student model fine-tuned on just 100K of these difficulty-pruned CoT samples outperforms a model distilled on 800K original Long CoT samples.<n>Our method also generalizes well: across 11 diverse benchmarks, the shorter difficulty-aware CoTs achieve equal or better accuracy than Long chains, using far fewer tokens.
  arXiv  Detail & Related papers  (2025-05-26T09:04:44Z)
• Thinkless: LLM Learns When to Think [57.857534644932194]
  Reasoning Language Models, capable of extended chain-of-thought reasoning, have demonstrated remarkable performance on tasks requiring complex logical inference.<n>We propose Thinkless, a learnable framework that empowers an LLM to adaptively select between short-form and long-form reasoning.<n>On several benchmarks such as Minerva Algebra, MATH-500, and GSM8K, Thinkless is able to reduce the usage of long-chain thinking by 50% - 90%.
  arXiv  Detail & Related papers  (2025-05-19T17:24:16Z)
• DAST: Difficulty-Adaptive Slow-Thinking for Large Reasoning Models [30.184895117009457]
  This paper introduces Difficulty-Adaptive Slow Thinking (DAST), a novel framework that enables models to autonomously adjust the length of Chain-of-Thought (CoT) based on problem difficulty.<n>Experiments on diverse datasets and model scales demonstrate that DAST effectively mitigates overthinking while preserving reasoning accuracy on complex problems.
  arXiv  Detail & Related papers  (2025-03-06T14:23:06Z)
• Dualformer: Controllable Fast and Slow Thinking by Learning with Randomized Reasoning Traces [40.127653552777204]
  We present dualformer, a single Transformer model that seamlessly integrates both the fast and slow reasoning modes.<n>At inference time, dualformer can be easily configured to execute in either fast or slow mode, or automatically decide which mode to engage.<n>It outperforms baselines in both performance and computational efficiency across all three modes.
  arXiv  Detail & Related papers  (2024-10-13T16:53:02Z)
• 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)
• Answering Unseen Questions With Smaller Language Models Using Rationale Generation and Dense Retrieval [9.136948771060895]
  We evaluate two methods for further improvement in this setting. Both focus on combining rationales generated by a larger Language Model with longer contexts created from a multi-hop dense retrieval system. Our single best Reasoning model materially improves upon strong comparable prior baselines for unseen evaluation datasets.
  arXiv  Detail & Related papers  (2023-08-09T05:06:39Z)

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.