- Rating
- 3(+1)
- Confidence
- 2(−1)
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 3(+1)
Read review
Paper summary
The paper investigates how answer-candidate composition reshapes layer-wise preference formation in decoder-only language models and whether candidate-dependent hidden states causally influence final predictions. It compares paired benchmark and unrelated-option inputs across six Llama and Qwen models, 15 MCQA tasks, followed by full-state activation-patching experiments on six tasks using panels of 36, 180, and 360 questions.
Critical strengths
A large sum of empirical evaluation. The paper evaluates six Llama and Qwen models across fifteen general and domain-specific tasks under different settings, resulting in a large collection of paired model-question records. The consistency of the reported patterns across models, tasks, and alternative readout settings provides broad empirical support for the main observations.
Critical weaknesses
-
The distinction from prior activation-patching work should be discussed in greater depth. The paper applies full-state activation patching at a selected token and layer and measures the resulting behavioral shift. Similar state-transfer interventions and the sensitivity of MCQA predictions to answer options have been studied previously. The paper should discuss deeper such as which model components implement the observed effect rather focusing the phenomenon.
-
Mechanistic expainasion beyond layer-wise observations should be discussed. The paper provides an empirical characterization of how answer preferences and intervention effects vary across layers. Prior information-flow studies have further localized such dynamics to specific attention heads, MLPs, and computational paths. A component- or path-level analysis would help explain why the identified layers influence the final decision.
Questions for authors
Please find the weakness above.
6 comments
Response for Reviewer HPZR [2/2]
Weakness 2: Which model components and information-flow pathways explain how layer-wise effects influence the final decision?
To move beyond layer-wise observations and identify specific computational components, we conducted a two-stage component screening followed by independent confirmation on Qwen3-4B-Base.
(1) Systematically identifying influential attention heads. Starting from the original intervention layer (Layer 9), we screened the attention and MLP outputs at the answer-prediction position across the subsequent 26 layers, evaluating 52 candidate components. The highest-ranked candidates were the attention output at Layer 24 and the MLP output at Layer 29. We then individually tested all 64 attention heads in these two layers using bidirectional restoration and blocking interventions. The screening score was defined as the smaller of the task-balanced mean restoration and blocking effects. Only heads with positive effects in both directions and output changes exceeding the numerical error bound were retained. Exactly three heads met these criteria:
| Discovery rank | Attention head (layer, head) | Restoration | Blocking | Screening score |
|---|---|---|---|---|
| 1 | (24, 30) | 0.05684 | 0.07186 | 0.05684 |
| 2 | (29, 0) | 0.00095 | 0.00798 | 0.00095 |
| 3 | (24, 0) | 0.00209 | 0.00038 | 0.00038 |
All layer and head indices are zero-based. We also performed head screening on Llama-3.2-3B, but no head met the criterion for confirmation.
(2) Confirming the causal contribution of the selected head group. We froze these three heads and jointly intervened on their outputs at the answer-prediction position using independent confirmation sets of 120 CSQA and 80 MMLU-General questions. Compared with norm-matched random interventions, the selected head group produced the following effects:
| Model | Task | Questions | Restoration advantage ↑ | Blocking advantage ↑ |
|---|---|---|---|---|
| Qwen3-4B-Base | CSQA | 120 | 0.06246 (0.00421) | 0.07543 (0.00801) |
| Qwen3-4B-Base | MMLU-General | 80 | 0.00808 (0.00154) | 0.01416 (0.00138) |
Note: Values represent mean (sample variance) of normalized score-gap changes beyond norm-matched random interventions. Both restoration and blocking passed Holm correction on CSQA, while blocking alone passed on MMLU-General.
(3) Identifying the primary individual contributor and examining information flow. We further performed individual interventions on the three selected heads. On the CSQA confirmation set, Layer 24, Head 30 showed the clearest individual contribution: its restoration and blocking effects were 0.06357 and 0.07482, respectively, close to the corresponding raw means of the three-head joint intervention (0.06418 and 0.07650). In comparison, the individual effects of the other two heads were close to zero.
| Attention head | Restoration ↑ | Blocking ↑ |
|---|---|---|
| Layer 24, Head 30 | 0.06357 (0.00415) | 0.07482 (0.00858) |
| Layer 29, Head 0 | −0.00015 (0.00046) | −0.00152 (0.00066) |
| Layer 24, Head 0 | 0.00152 (0.00055) | 0.00046 (0.00080) |
Furthermore, Key-Value (K/V) interventions showed that jointly exchanging K/V information across all context positions accessible to the answer-prediction token changed the score gap in both restoration and blocking directions. Together, these findings connect systematic component screening, independent head-level confirmation, individual head contributions, and contextual information-flow interventions, providing a more detailed mechanistic explanation of how distractor-dependent representations influence final answer preferences. We will incorporate these analyses into the revised manuscript.
Follow-up to Reviewer HPZR
Thank you for your careful review and constructive questions. We would greatly appreciate a brief reply on whether our rebuttal addresses your concerns.
We clarified how our study differs from prior activation-patching work, traced part of the distractor effect to early model components, and identified a specific attention head in Qwen3-4B-Base that influences answer preferences on CommonsenseQA.
Thank you again for your time and consideration.
Response for Reviewer HPZR [1/2]
Weakness 1: How does this work differ from prior activation-patching studies, and which components contribute to the observed effect?
Our work addresses a distinct question from prior activation-patching studies: how does changing only distractor content reshape a model's preference for the same correct answer? We construct paired inputs that keep the question stem, correct answer and its position, and option count fixed while replacing only incorrect options. We track the emergence of sustained correct-answer preference across layers and use the same strongest incorrect option as a fixed competitor in our interventions. Across 90 model–task settings, benchmark distractors delayed the last preference revision in 86 settings and shortened the persistent correct-preference span. Full-state transfer on an independent 180-question set improved mean accuracy by 8.78 percentage points. Its effect on the correct-versus-fixed-distractor score gap also exceeded both norm-matched random and cross-question controls.
(1) Identifying the state components contributing to the effect. We conducted component-removal experiments on Llama-3.2-3B and Qwen3-4B-Base using MMLU (80 questions) and CSQA (120 questions). We separately removed the state difference at the intervention-layer input, the attention output difference, and the MLP output difference, followed by downstream recomputation. Removing the input-state difference produced the largest mean attenuation in all four model–task settings. Removing the MLP output difference also significantly attenuated the transfer effect on CSQA for both models after Holm correction, whereas removing the attention output difference yielded no significant effect.
(2) Tracing the upstream sources of state differences. We identified upstream sources on an independent discovery set and evaluated them on held-out questions by swapping module outputs between paired inputs, recomputing the state at the original intervention site, and repeating the transfer. Introducing unrelated-distractor outputs partially restored the transfer gain, whereas the reverse swap attenuated it. The identified sources were input embeddings and Block 0 MLP for Llama-3.2-3B, and Block 0 attention and MLP for Qwen3-4B-Base. All 16 restoration and blocking tests passed Holm correction.
Note: All entries report mean ΔG (sample variance), where G is the normalized score gap between the correct answer and the fixed distractor. In the first table, larger values indicate greater attenuation of the full-state transfer effect after component removal. In the second table, restoration measures the gain over the original-question baseline, while blocking measures the loss relative to full-state transfer. Variances are calculated across questions after averaging option rotations within each question.
Together, these results connect the distractor-dependent state differences to specific upstream modules and identify the residual and MLP contributions underlying their downstream effects. We will incorporate these analyses and clarify the distinction from prior activation-patching work in the revised manuscript.