finding
active
finding:claude-models-use-hints-24-more-frequently-than-deepseek-models-relativeClaude models use hints 24% more frequently than DeepSeek models (relative)
Consistent for correct and incorrect hints.
Source paper
extracted_from(2025) · Yanda Chen · Joe Benton · Ansh Radhakrishnan · Jonathan Uesato +11
Neighborhood — ranked by edge-count
Papers (1)
paper
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Reasoning > non-reasoning.
- CoTs may hide problematic reasoning.
- Average faithfulness 3.6x higher for Claude series and 1.7x for DeepSeek series (reasoning vs non-reasoning)finding0.773Quantifies reasoning advantage.
- All four models change answer to hint answer significantly more often than to a non-hint answer on MMLU/GPQAfinding0.767Validates metric assumption.
- Claude-instant-1.2 is the most accurate (91.1%) and most coherent (88.6%) LM on the Leap-of-Thought dataset.finding0.767Main result from Experiment 2, Table 2.
- Claude 3.7 Sonnet verbalizes used hints at least 1% in 5/6 settings vs Claude 3.5 Sonnet 4/6finding0.763Reasoning > non-reasoning.
- Trend observed in Experiment 2 results.
- Base failure mode.