finding
active
finding:mixed-lr-probe-achieves-0-386-to-0-826-accuracy-against-6-7-chance-but-fails-on-natural-documentsMixed LR probe achieves 0.386 to 0.826 accuracy against 6.7% chance, but fails on natural documents
No evidence of chronic emotion representation.
Source paper
extracted_from(2026) · Nicholas Sofroniew · Isaac Kauvar · William Saunders · Runjin Chen +12
Neighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (1)
hypothesis
- Explains negative probing results.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Likely-trained MM probe is a surprisingly effective causal baseline due to correlation between truth and probability on sp_en_trans
- Probe achieves selectivity of 4.20 on pythia-410m, slightly exceeding DAS selectivity of 3.96finding0.779Key result showing that for models larger than pythia-70m, probe selectivity matches or exceeds DAS selectivity
- Linear probe achieves 100% classification accuracy for almost all components in Pythia-6.9B gender taskfinding0.771Demonstrates that linear probes can overestimate causal relevance; probes succeed on non-causally-relevant representations
- Core negative result: the binary detection paradigm cannot distinguish genuine introspection from uniform output bias
- At n=256 best-of-n, probe scoring outperforms Gemma LLM-judge scoring by nearly 15 percentage pointsfinding0.767Test-time scaling comparison.
- Larger models linearly represent more general concepts including truth
- Striking cross-domain generalization result supporting the claim that larger models represent abstract truth
- Intrinsic noise in representation.