- Rating
- 1
- Confidence
- 2
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
CSS-Bench tests whether LLM choices change when action labels are swapped while payoffs stay fixed. The paper finds label sensitivity across nine models and investigates it with probes, activation patching, and steering. Its main mechanistic claim is that label information becomes decodable early but affects decisions much later.
Critical strengths
-
The paired design is useful. Both prompts use the same labels and tokens, making logit comparisons and activation interventions straightforward.
-
The authors report several limiting results, including sign reversals across templates, poor steering transfer to unseen labels, and failed geometric localization. This helps readers see where the evidence stops.
Critical weaknesses
-
My first concern is whether these models can already solve the payoff task. Averaging the two label assignments in Table 2 gives 43.0% accuracy, with 31 of 36 cells below 50%. There is no base accuracy with labels such as A/B, and the correct option in the Prisoner's Dilemma is always Slot B. Simply changing order moves Pythia-410M from 27.8% to 54.6% (Table 16). Label effects could therefore include strong position or word preferences; they do not by themselves establish semantic interference with payoff-based decisions. Neutral labels also have the largest average effect, which weakens the semantic framing. The word pools are tiny: moral has only 4/2 words (Table 4), reused across 3,000 pairs per cell. Instance-level significance does not establish generalization across words. Semantically empty labels, balanced positions, and per-word-pair results are needed to tell what the benchmark measures.
-
The mechanism analysis leaves straightforward alternative explanations open. Both probes train only on canonical prompts, where the payoff-optimal slot and positive-word slot are identical (Section 6.1, Appendix H). Reading word position is enough to solve the probe task; this does not identify payoff information. Patching only replaces the final position, leaving label information elsewhere for later layers to reread, so late recovery need not be specific to the label mechanism. The 0.25 and 0.76 depths measure different things and do not establish a causal delay. Steering has a related issue: the correct counterfactual answer is always the negative word, so favoring that word class can raise accuracy. Unseen-label accuracy is only 44%. Decorrelation in probe training, multi-position and payoff-only patching controls, and steering on canonical prompts are needed to rule out these explanations.
Questions for authors
-
What accuracy do the models reach with A/B-style labels and balanced correct-option positions? Please break this down by game and report effects by word pair. This would directly affect my reading of the semantic-sensitivity claim.
-
What are the probes' absolute early-layer accuracies? Can they decode the payoff-optimal slot when trained on both canonical and counterfactual prompts?
-
Does recovery stay late when label positions or multiple positions are patched layer by layer? Does changing only payoffs produce a similar curve?
-
What happens when the same steering direction is applied to canonical prompts? Why does Qwen2.5-7B moral score 50.5%/48.3% in Table 2 but 3.2%/95.8% in Table 13? Is this due to the word split? Please also check the direction sign in Eq. 10.
Minor: check the sign conventions in Tables 2 and 15 and Appendix M. Some claims in Appendix A lack citations, and Evans et al. is missing from the references.