finding
active
finding:steering-had-no-clear-effect-on-behavior-for-some-llms-e-g-phi-4-mini-instructSteering had no clear effect on behavior for some LLMs (e.g. phi-4-mini-instruct)
Limitation.
Source paper
extracted_from(2025) · Eric Bigelow · Daniel Wurgaft · YingQiao Wang · Noah D. Goodman +3
Neighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (1)
hypothesis
- Explanation for phi-4 null effect.
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.
- Empirical comparison showing advantage of SAE features in low-data regime.
- At steering strength 1.4 all factual claims are incorrect yet the LLM evaluator still gives a high scorefinding0.799Pitfall of LLM-based steering evaluation.
- The paper's critique of the standard linear steering baseline, supported by the days-of-week demo.
- Core empirical claim comparing steering approaches on cyclic concepts.
- Empirical result demonstrating the failure mode of linear steering when concept geometry is cyclic.
- Validates using chain-of-thought belief monitoring as proxy for behavioral steering efficacy.
- Mechanism claim supported by transcript analysis and the fact that the steering vector was extracted from a model that never writes type hints.
- DIM steering has less off-target effect but smaller maximum on-target effect than linear probe steeringfinding0.787Method tradeoff.