Entity · paper

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

paperactiveprovisionaltoken-budget-saturation-and-mechanistic-early-detection-of-reasoning-non-convergence-in-chain-of-thought-models-19653835·1 events·first seen 27h ago

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

Co-occurring entities

AIME DeepSeek-R1-Distill-Qwen

More like this (12)

When the Chain of Thought Knows Better: Failure Modes in Multi-Turn Reasoning Models Beyond the Commitment Boundary: Probing Epiphenomenal Chain-of-Thought in Large Reasoning Models Look Light, Think Heavy: What Multimodal Chain-of-Thought Reasoning Can and Cannot Do Agentic Chain-of-Thought Steering for Efficient and Controllable LLM Reasoning Chain-of-Thought Reasoning Reasoning Language Models Does Reasoning Preserve Alignment? On the Trustworthiness of Large Reasoning Models What Makes Effective Supervision in Latent Chain-of-Thought: An Information-Theoretic Analysis Long-context Reasoning Benchmarks Predicting Future Behaviors in Reasoning Models Enables Better Steering Bridging the Gap Between Latent and Explicit Reasoning with Looped Transformers chain-of-thought monitoring

Recent events (1)

5arXiv · cs.CL·27h ago·source ↗

Linear probes on chain-of-thought hidden states can predict reasoning non-convergence early

A new arXiv preprint studies the bimodal convergence pattern in chain-of-thought models like DeepSeek-R1-Distill-Qwen-7B, where generations either complete within a token budget (90.3% accuracy on AIME) or exhaust it without concluding (6.6% accuracy), with a 62% overall convergence rate. The authors train linear probes on hidden-state activations at early token positions (50-300) and find that layer-20 activations at token 150 achieve AUC 0.608, reliably above chance and outperforming behavioral baselines from token entropy and repetition statistics. The results suggest convergence fate is partially encoded in intermediate representations early in generation, pointing toward early-exit inference and adaptive compute allocation strategies. Statistical evidence is modest (permutation test p=0.063), limiting strong conclusions.

Evaluation and Benchmarking Inference Economics AIME DeepSeek-R1-Distill-Qwen Token Budget Saturation and Mechanistic Early Detection of Reasoning Non-Convergence in Chain-of-Thought Models