On GPQA-Diamond multihop questions, activation probes show genuine belief shifts during CoT generation rather than early stabilization, contrasting with MMLU

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

• Activation probing detects final answer belief earlier in CoT than CoT monitor on both models, with especially pronounced gap on easy MMLU questionsfinding0.830
  Comparative finding establishing activation probing as superior to text-level monitoring for early belief detection
• can activation probing enable efficient adaptive computation by detecting when a model's belief has stabilized during CoT generation?question0.805
  Practical question addressed by the probe-guided early exit experiments
• Probe-guided early exit reduces tokens by up to 30% on GPQA-Diamond with similar accuracy on DeepSeek-R1 671B and GPT-OSS 120Bfinding0.785
  Quantitative efficiency result on hard benchmark, smaller reduction reflecting genuine reasoning need
• Probe-guided early exit reduces tokens by up to 80% on MMLU and 30% on GPQA-Diamond with similar accuracyclaim0.784
  Practical efficiency claim for using activation probes to enable adaptive computation
• Model final answer is decodable from activations far earlier in CoT than CoT monitor detects on MMLU recall-based questions for both DeepSeek-R1 671B and GPT-OSS 120Bfinding0.774
  Core empirical result demonstrating early belief formation in easy tasks
• We hypothesize that degraded generalization on benchmarks like MMLU may reflect the computational demands of the tasks.hypothesis0.774
  Connecting the paper's task-difficulty findings to prior observations of weak generalization on complex QA benchmarks.
• Multimodal-CoT with vision features achieves higher validation accuracy at early training epochs (epoch 1-3) compared to one-stage and two-stage language-only baselines on ScienceQAfinding0.773
  Evidence that multimodal information accelerates convergence speed during training.
• Direct probes over learned activations in standard basis may fail to reveal the actual causal role of representations because they are highly distributedclaim0.767
  Supported by the finding that non-trivial rotations are required to find aligned representations.