Unverified50% confidenceBenchmarkExact time
On Llama 3.2 3B, the marginal improvement on the ConfabQA benchmark was within the range of fold-to-fold fluctuation, but refusal-specific probes achieved 18–25 percentage point improvements on the PopQA and TriviaQA datasets.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/3/2026
First Seen
Valid until: 2/1/2027
Sources
Related Entities
Related Claims
UnverifiedThe correctness signal is model-dependent and transferable only on Llama 3.2 3B: when probe margins were transferred among ConfabQA, PopQA and TriviaQA, 81% of the margin was preserved on the test majority class.75% similarUnverifiedOn the Qwen3-1.7B and Gemma 2 2B models, linear probes significantly outperformed prompt-feature-based methods in distinguishing imminent refusals from confabulations, improving by 7.4 and 16.3 percentage points respectively.73% similarUnverifiedAfter re-evaluating with non-overlapping anchors, the net advantage instead grew from +4.1pp to +5.5pp, but significance dropped from 2.4 standard errors to 1.4 standard errors; the inflation factor is about 2.4x, exactly equal to sqrt(912/163).68% similarUnverifiedAfter expanding the context from 256 tokens to 2,048 tokens, the evidence gain under fixed-answer conditions on FEVEROUS improved by 3.84 (Qwen) and 3.10 (Llama) percentage points respectively.67% similarUnverifiedVisKG-LM achieves improvements of 1.2, 0.8, and 4.3 points over the GreaseLM baseline on the three benchmarks CommonsenseQA, OpenBookQA, and MedQA-USMLE, respectively.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1128933API
curl https://kongchang.com/api/v1/knowledge/claims/1128933MCP
get_claim(id=1128933)