Unverified50% confidenceBenchmarkExact time
In Bengali, Gemma 2 9B's stereotype-aligned bias increases, while Mistral-Nemo's bias decreases.
1
Sources
50%
Confidence
Long-term
Relevance
11/10/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedLaya scored 0% on Khmer text yet reported an average confidence of 95.2%, a textbook case of calibration failure.53% similarUnverifiedIn disambiguated contexts, the accuracy improvement is +36.59 percentage points for Bengali and +17.71 percentage points for English.52% similarUnverified"Cheap, Consistent, Calibrated, but Biased?" tested 8 typed decision models (TDMs) with probabilistic outputs and 7 LLM evaluators, finding that each TDM was biased on 3-8 of 12 bias tests, with emotional tone having the largest effect, causing accuracy to plummet from 99% to 42%.51% similarUnverifiedThe largest uncorrected effect was that shifting language range away from neutral paraphrase changed other players' votes by about 4 percentage points, but this effect size is not statistically robust.50% similarUnverifiedFew-shot reasoning significantly reduced the hallucination rate: the strongest model produced unsupported QoS requirements in about 1% of responses, whereas performance under the zero-shot setting was markedly worse.50% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1154786API
curl https://kongchang.com/api/v1/knowledge/claims/1154786MCP
get_claim(id=1154786)