Unverified50% confidenceBenchmarkExact time
On 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.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/3/2026
First Seen
Valid until: 2/1/2027
Sources
Related Entities
Related Claims
UnverifiedOn the Qwen-3.5-4B-Instruct model, the top two principal axes matched the style dimensions of human spontaneous requests with a precision of 72.8% and a macro-averaged recall of 43.6%.73% similarUnverifiedOn 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.73% similarUnverifiedIn matched-frame controlled experiments conducted on six Qwen-series models, the caption approach improved average accuracy by 3.22 points and 2.55 points respectively under frame-aligned conditions.72% similarUnverifiedOn held-out Qwen3-32B dialogue segments, with the threshold set to flag 5% of benign calls, the linear probe caught 82% of leaking calls.72% similarUnverifiedThe fine-tuned Qwen3.5-4B model improved accuracy on BU Bench V1 from 15% to 53%.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1128932API
curl https://kongchang.com/api/v1/knowledge/claims/1128932MCP
get_claim(id=1128932)