Unverified50% confidenceFactExact time
As test-time compute investment increases, the divergence between the model's answers and human answers also widens; the original text describes this phenomenon as 'may be a general feature of stronger language models' and points out that it introduces 'a potential general conflict between capability and alignment'.
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
First Seen
Sources
Related Entities
Related Claims
Unverified自然语言驱动的测试存在确定性问题，AI 对同一描述的理解可能存在波动，影响结果可重复性78% similarUnverifiedThe research conclusions challenge the common assumption that the advantages of multilingual models stem from blurring language boundaries and emphasizing commonalities, suggesting that moderately emphasizing language differences is in fact a superior strategy.75% similarUnverifiedOOD detection for autoregressive language models is more difficult than for traditional machine learning because the linguistic compositional space is vast, semantic OOD does not correspond to surface-form OOD, and adversaries can construct inputs that appear statistically in-distribution but semantically trigger OOD behavior.74% similarUnverifiedNYU's Gary Marcus believes that the emergent capabilities of large language models may reflect threshold effects in evaluation metrics more than a qualitative change in the models' cognitive abilities.74% similarUnverifiedVocalGrad serves as a reminder: high scores of multimodal models on semantic tasks may mask gaps in perceptual ability, and when models rely on textual priors or linguistic reasoning to guess answers, benchmarks will overestimate their true ability.74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1129300API
curl https://kongchang.com/api/v1/knowledge/claims/1129300MCP
get_claim(id=1129300)