Unverified50% confidenceFactExact time
The 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.
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOn 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.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.67% similarUnverifiedThe random-position null baseline for Qwen3-4B was only 11.3%, and the upper bound for directly patching lower-level prompt facts was 83%.64% similarUnverifiedAmong the three tested models, only Qwen2.5-3B can reliably refuse to answer when the evidence does not support an answer.64% 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.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1128939API
curl https://kongchang.com/api/v1/knowledge/claims/1128939MCP
get_claim(id=1128939)