Probe-guided early exit reduces tokens by up to 80% on MMLU and 30% on GPQA-Diamond with similar accuracy

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• On GPQA-Diamond multihop questions, activation probes show genuine belief shifts during CoT generation rather than early stabilization, contrasting with MMLUfinding0.784
  Empirical finding contrasting difficult questions with easy ones, supporting genuine reasoning on hard tasks
• Probe-Guided Early Exitconcept0.778
  Using activation probes to terminate CoT generation early when the model's belief is already stable, saving compute
• QwQ-32B accuracy on MMLU Formal Logic stays between 95.5% and 96.3% across all intervention strengths while tokens reduced from 1716.6 to 1481.4 at -0.96finding0.741
  Demonstrates reflection redundancy in larger models on non-mathematical reasoning
• Up to 33.6% reasoning tokens saved on MMLU subsets with stepwise steering while maintaining accuracy in larger modelsfinding0.741
  Maximum token savings achieved by ReflCtrl on non-mathematical general reasoning tasks
• Activation probing detects final answer belief earlier in CoT than CoT monitor on both models, with especially pronounced gap on easy MMLU questionsfinding0.733
  Comparative finding establishing activation probing as superior to text-level monitoring for early belief detection
• For simple factual tasks F0-F3, probe directions show a sharp geometric transition in middle layers, with late-layer probes converging to high cosine similarity; A3 and F4-F5 show no clear transition.finding0.728
  Geometric evidence for convergence to stable truth directions only for simpler tasks.
• MM probe trained on likely dataset achieves NIE of 0.70 (false→true) on LLaMA-2-13B, surprisingly strong but weaker than truth probesfinding0.726
  Likely-trained MM probe is a surprisingly effective causal baseline due to correlation between truth and probability on sp_en_trans
• Probe-based ranking reduces harmful behavior by 63% via datapoint filteringfinding0.724
  Primary quantitative result: probe method outperforms gradient-based and LLM-judge alternatives at lower computational cost.