Token Budget Saturation and Mechanistic Early Detection of Reasoning Non-Convergence in Chain-of-Thought Models

• URL: http://arxiv.org/abs/2607.21433v1
• Date: Thu, 23 Jul 2026 15:37:04 GMT
• Title: Token Budget Saturation and Mechanistic Early Detection of Reasoning Non-Convergence in Chain-of-Thought Models
• Abstract summary: Chain-of-thought reasoning models exhibit a bimodal convergence pattern.<n>Generations terminate within a token budget (converged) or exhaust it without reaching a conclusion (non-converged)<n>We show that converged generations achieve 90.3% accuracy on AIME 1983-2024 while non-converged ones achieve only 6.6%, with an overall convergence rate of 62.0%.
• Score: 1.4563514298965277
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Chain-of-thought reasoning models such as DeepSeek-R1-Distill-Qwen-7B exhibit a bimodal convergence pattern: generations either terminate within a token budget (converged) or exhaust it without reaching a conclusion (non-converged). We characterize this phenomenon empirically, showing that converged generations achieve 90.3% accuracy on AIME 1983-2024 while non-converged ones achieve only 6.6%, with an overall convergence rate of 62.0%. We then ask whether this outcome is detectable early in the thinking chain using internal model representations. Training linear probes on hidden-state activations at token positions 50-300, we find that layer-20 activations at token 150 achieve AUC 0.608 (+-0.080, 5-fold CV), reliably above chance even at token 50. Activation probes consistently outperform behavioral baselines derived from token entropy and repetition statistics. A sweep-level permutation test yields p=0.063 (100,000 permutations), consistent with a modest signal that our sample size cannot confirm at conventional thresholds. These findings suggest that convergence fate is partially encoded in intermediate representations well before the generation ends, opening a path toward early-exit inference and adaptive compute allocation.

Related papers

• Coarse composition suffices: tabular in-context learning for multi-activity antimicrobial peptide profiling [2.571440349338848]
  We show that a simple, sequence-only pipeline can match and surpass multimodal, structure-conditioned deep models.<n>On ESCAPE (82,359 peptides; five labels), a label-powerset TabPFN model achieves mAP-5 = 77.8%, improving on the previously best reported 72.1%.
  arXiv  Detail & Related papers  (2026-08-31T06:50:09Z)
• GradCuit: Credit-Assigned Gradient Flow Enables Robust and Interpretable Test-Time Latent Reasoning [61.67411833252235]
  We introduce GradCuit, which inserts optimizable latent states at a selected Transformer layer between the hidden representations of the prompt and the generated continuation.<n>Across five instruction-tuned backbones, three reasoning benchmarks, and two answer formats, GradCuit achieves an average accuracy of 64.5%.<n>For interpretability, token-level gradient attribution reveals that latent influence concentrates on reasoning-connector tokens, while layer analysis identifies early-to-middle Transformer layers as the most effective optimization space.
  arXiv  Detail & Related papers  (2026-08-03T17:55:24Z)
• SketchMamba: A Lightweight State-Space Model for Joint Progressive Sketch Classification and Stroke Auto-Completion [0.0]
  Existing vector-sketch models treat recognition and generation as separate tasks.<n>We present SketchMamba, a single causal sequence model that continuously classifies a sketch from any partial prefix while simultaneously generating its continuation.
  arXiv  Detail & Related papers  (2026-07-26T10:05:23Z)
• CUSUM-Shaped Inference-Time Monitoring and Targeted Re-Decoding for Quantized Small Language Model Reasoning [0.6999740786886536]
  MGT-B maps overlapping windows of pre-sampling uncertainty and degeneration features to position-conditional empirical tail probabilities.<n>We audit whether the effect persists on problem identities first observed after the manual choice of log threshold h = 10.<n>Neither analysis is confirmatory, and the empirical factors are not established as a valid e-process or e-detector.
  arXiv  Detail & Related papers  (2026-07-22T13:34:17Z)
• When Thinking Hurts: Epistemic Signals in the Reasoning Chains of Visual Language Models [0.0]
  We provide the first three-family empirical characterisation of answer entropy behaviour in visual language models (VLMs)<n>We find three qualitatively distinct patterns: Qwen3-VL-8B-Thinking shows complete collapse; GLM-4.1V-9B-Thinking shows no collapse; and InternVL3-8B shows selective thinking.<n>Across all three thinking-mode models, thinking chain entropy outperforms answer entropy on the subset where chains are generated.
  arXiv  Detail & Related papers  (2026-07-09T02:19:16Z)
• Automated Proving of Shannon-Type Entropy Inequalities via Fine-Tuned Language Models and Guided Tree Search [50.16356451328644]
  Proving Shannon-type entropy inequalities is a fundamental task in information theory.<n>We investigate whether small-scale large language models, fine-tuned on atomic proof steps, can automate this process.<n>We show that GPT-5.5 solves 1.7% samples under zero-shot prompting while Psitip solves 33.3% samples.
  arXiv  Detail & Related papers  (2026-06-04T05:43:12Z)
• Unveiling the Entropy Dynamics of Chain-of-Thought Reasoning [81.57028614960576]
  We show that the Confidence Region possesses two critical properties: 1) High Reliability -- answers in the confidence region become highly accurate and stable, and 2) High Redundancy -- models generate unnecessary tokens long after reaching the correct answer.<n>These properties unlock more efficient and reliable inference strategies.
  arXiv  Detail & Related papers  (2026-06-01T10:11:14Z)
• ESPO: Early-Stopping Proximal Policy Optimization [78.79610718910628]
  ESPO (Early-Stopping Proximal Policy Optimization) detects trajectory failure on-the-fly and terminates rollouts early.<n>On DeepSeek-R1-Distill-Qwen-7B trained for mathematical reasoning, ESPO surpasses PPO on AIME2024 (46.28% vs. 45.25%), AMC2023 (85.83% vs. 82.94%), and MATH-500 (87.42% vs. 85.43%)
  arXiv  Detail & Related papers  (2026-05-28T12:40:22Z)
• Tracing Uncertainty in Language Model "Reasoning" [46.61902399979181]
  Language model (LM) "reasoning", commonly described as Chain-of-Thought or test-time scaling, often improves benchmark performance.<n>We study these dynamics by treating the "reasoning" traces, the intermediate token sequences generated by LMs, as evolving model states.
  arXiv  Detail & Related papers  (2026-05-08T14:16:37Z)
• Relational Preference Encoding in Looped Transformer Internal States [0.0]
  We investigate how looped transformers encode human preference in their internal iteration states.<n>Using Ouro-2.6B-Thinking, a 2.6B- parameter looped transformer with iterative refinement, we extract hidden states from each loop iteration.<n>We train lightweight evaluator heads to predict human preference on the Anthropic HH-RLHF dataset.<n>Our pairwise evaluator achieves 95.2% test accuracy on 8,552 unseen examples, surpassing a full-batch L-BFGS probe (84.5%) while the base model remains completely frozen.
  arXiv  Detail & Related papers  (2026-04-10T20:00:49Z)
• Unregularized Linear Convergence in Zero-Sum Game from Preference Feedback [50.89125374999765]
  We provide the first convergence guarantee for Optimistic Multiplicative Weights Update ($mathtOMWU$) in NLHF.<n>Our analysis identifies a novel marginal convergence behavior, where the probability of rarely played actions grows exponentially from exponentially small values.
  arXiv  Detail & Related papers  (2025-12-31T12:08:29Z)
• LYNX: Learning Dynamic Exits for Confidence-Controlled Reasoning [15.597220136913258]
  LYNX is an online early-exit mechanism that turns a model's own hidden-state awareness into confidence-controlled stopping decisions.<n>We train and calibrate this probe once on a generic mathematical corpus and reuse it unchanged across benchmarks, decoding temperatures, and even non-mathematical tasks.
  arXiv  Detail & Related papers  (2025-12-05T00:04:42Z)
• Inference-Time Chain-of-Thought Pruning with Latent Informativeness Signals [6.5422130090856925]
  Self-Truncation Best-of-N (ST-BoN) mitigates this by truncating unpromising paths early.<n>We present KL-Adjusted Pruned Path Algorithm (KAPPA), an inference-time method that combines Kullback-Leibler divergence, confidence, and entropy into a principled scoring function.
  arXiv  Detail & Related papers  (2025-11-01T20:41:22Z)
• Semi-supervised Contrastive Learning with Similarity Co-calibration [72.38187308270135]
  We propose a novel training strategy, termed as Semi-supervised Contrastive Learning (SsCL) SsCL combines the well-known contrastive loss in self-supervised learning with the cross entropy loss in semi-supervised learning. We show that SsCL produces more discriminative representation and is beneficial to few shot learning.
  arXiv  Detail & Related papers  (2021-05-16T09:13:56Z)

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.