acceptodds
Withdrawn from ICLR 2027

Position Still Matters: Localizing and Controlling Latent Positional Effects in LLM Comparisons

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language models are increasingly used as evaluators, yet their judgments can depend on presentation order rather than only on relevant evidence. Evaluating both presentation orders can reveal inconsistent choices, but it does not show whether order still influences the decision when the final answer remains unchanged, where that influence enters the computation, or whether it can be reduced without weakening the comparison itself. We study these questions through a three-stage mechanistic framework. First, we decompose the model's decision margin into separate effects of comparison evidence and presentation position, revealing substantial positional shifts even in models whose choices are nearly invariant to swapping. Second, activation patching localizes this influence to late prompt states, while head-level interventions characterize how individual attention heads affect positional dependence and evidence sensitivity. Third, we construct interventions that reduce the positional effect under explicit constraints on comparison performance. On held-out numerical comparisons, the procedure removes 60–96% of the positional effect in five of six models, improving held-out balanced accuracy by up to 20 percentage points; in the remaining model, the same constraints permit only limited control. Reapplying the same procedure to country-population comparison and letter counting reduces the positional effect in all 12 model–task settings, by 21–100%, whereas the intervention selected on numerical comparison transfers inconsistently. These results show that behavioral order invariance can mask substantial latent positional influence, and that this influence can be causally localized and selectively controlled inside the model.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, position bias, pairwise comparison, activation patching, causal intervention, LLM evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_MRzsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents an interesting finding that positional bias can persist even when LLM decisions remain unchanged, and proposes a framework to localize and mitigate it while preserving evidence sensitivity. Their integration provides useful empirical insights. Further analysis of the mean-centered intervention and fallback cases could strengthen the mechanistic understanding of why positional bias is easier to control in some models than others.

Critical strengths
  • The motivation is clear. The paper identifies an important distinction between behavioral consistency and latent positional bias, showing that substantial positional effects can persist even when LLM predictions remain unchanged.
  • The empirical part is systematic. The paper integrates margin-based measurement, activation patching, and constrained attention-head interventions into a coherent framework. Experiments across six models and multiple comparison tasks demonstrate substantial bias reduction while largely preserving evidence sensitivity.
Critical weaknesses
  • The proposed framework primarily combines established techniques, including logit-margin regression, activation patching, and attention-head scaling.
  • The mean-centered intervention scales entire attention-head outputs without explicitly separating positional information from task-relevant evidence. Consequently, it remains unclear whether the observed bias reduction reflects selective control of positional representations or compensating changes across attention heads. Comparisons with alternative intervention strategies would strengthen the mechanistic claims.
  • The fallback procedure relies on heuristic head selection and performance thresholds, with limited explanation of why some models permit near-complete bias removal while others do not.
Questions for authors
  • The intervention parameterization in Eq. (4) appears inconsistent with the attenuation procedure, since α = 0 corresponds to no intervention. The authors should clarify the implementation and verify the effects across different gains.

  • Comparing mean-centered scaling with direct head scaling, mean ablation, and position-specific interventions under matched performance constraints would help establish whether the proposed method selectively controls positional bias.

  • The fallback cases deserve further investigation through bias-evidence trade-off curves, particularly for Phi-3 and Mistral-7B, to clarify whether limited controllability arises from performance constraints or the restricted intervention search.

  • The stability of head selection and gain calibration should be evaluated across multiple development seeds, including the consistency of selected heads, held-out bias reduction, and evidence preservation.

  • The authors could examine conditional positional effects across different score gaps, τp(g)=E[m(1)−m(0)∣Δs=g]\tau_p(g)=\mathbb{E}[m(1)-m(0)\mid\Delta s=g], to determine whether reductions in average positional bias conceal substantial effects at specific comparison difficulties.

Reviewer_9CAJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  • Problem: when an LLM compares two items (A vs. B), the order in which the items are shown can push its decision. Usually people only check whether the final answer flips when the order is swapped. The paper argues this misses "hidden" order effects that shift the model's internal preference without flipping the answer.
  • Measure: the model is asked which of two single-token IDs has the higher numeric score. The authors read the logit difference between the two IDs at the answer position ("decision margin") and fit a simple linear model: margin = constant + beta_e * (score gap) + beta_p * (is the correct item in slot 2). beta_p is the positional effect, beta_e the "evidence sensitivity".
  • Finding 1 (Table 1): all six small open models (Gemma-2 2B/9B, Qwen2.5 3B/7B, Phi-3-mini, Mistral-7B) have a significant positional effect of 0.83 to 6.46 logits. Gemma-2-9B and Qwen2.5-7B are 99.8–100% accurate in both orders, yet still carry shifts of −2.57 and −2.73 logits.
  • Localize: activation patching shows the order effect enters the decision at two late token positions (end of the user message, then the answer position) in all six models (Fig. 1, Fig. 2). A per-head "census" measures how attenuating/amplifying each attention head changes beta_p and beta_e.
  • Control: heads are selected on a development set so that beta_p goes to zero while beta_e and the mean margin stay at >= 90% of baseline; the head set and gain are frozen and tested on held-out pairs. Result (Table 2): 60–96% of the positional effect removed in 5 of 6 models; Mistral-7B only 20%. Behavioral effect (Table 3): +3.8 points disfavored-order accuracy for Gemma-2-2B and Qwen2.5-3B, +20.3 balanced accuracy for Phi-3 (which starts at 71%), ~0 for the two ceiling models.
  • Transfer (Table 4): the frozen numerical-task intervention transfers badly to country-population comparison and letter counting (−300% to 78% removal, sometimes making bias worse). Re-running the selection procedure per task reduces the effect in all 12 model–task settings by 21–100%.
  • Validation: six models, held-out evaluation with bootstrap CIs, off-task cost measured as Wikipedia loss increase (0.25–2.56%). Code "will be released".
Critical strengths
  • Clean, well-controlled measurement and a disciplined selection/held-out protocol. The paired design isolates order, the decomposition is simple and interpretable, interventions are frozen before evaluation, and the authors even disclose that the fallback rule was written after seeing initial results and evaluate it only on fresh draws. This is why I trust the numbers that are reported, which matters for an interpretability paper whose value rests on credibility.
  • The transfer analysis is honest and informative: showing that a frozen head recipe can increase bias on a new task (e.g., Phi-3 letter counting, −300%) while the procedure still works is a genuinely useful negative result for anyone hoping for a one-shot "debias heads" fix. This separation of procedure generality from recipe transferability is the most novel conceptual contribution.
Critical weaknesses
  • The setting is a toy proxy for the motivating application and no baseline mitigation is compared. The intro and title frame the work around LLM evaluators/judges, but every experiment compares two single-token IDs with numeric scores, two countries, or two letter counts; there is no experiment on actual pairwise judging of generated text. Moreover, the control stage is never compared against existing methods the paper itself cites (PINE, Yu et al.'s single-dimension scaling, PriDe-style prior calibration, or simply averaging over both orders), nor against a trivial logit-offset baseline (subtracting the fitted beta_p). Without these, it is impossible to judge whether the head-level machinery buys anything over cheaper fixes, and the practical story (needs white-box access, per-task dev data, and re-selection for every task) is weak. This is the main reason I lean reject.
  • The behavioral payoff is small except for one degenerate case. The "up to 20 percentage points" claim comes only from Phi-3, which starts at 50.0% accuracy in one order with a mean margin of 1.22 logits and a non-significant beta_e of 0.001; for that model the intervention mostly inflates the margin (+150%). For the other five models, balanced accuracy changes by −0.2 to +3.8 points (Table 3). Relatedly, Gemma-2-2B's positional effect varies a lot across draws (delta = 0.30 in App. B, beta_p = 0.83 in Table 1, 1.14 in Table 2), which raises questions about how stable the small effects and the selected heads are.
Questions for authors
  1. Which data draw does Table 1 use, and why do its baseline beta_p values differ from the Table 2 baselines (e.g., Gemma-2-2B 0.83 vs. 1.14, Qwen2.5-3B −4.31 vs. −3.98)? Please report a seed/draw-level variance of beta_p and of the selected head sets.
  2. Please add at least one comparison with an existing mitigation (PINE, Yu et al. 2024, PriDe calibration, or two-order averaging) and a trivial baseline that subtracts the fitted beta_p from the margin, on the same held-out pairs, reporting beta_p removal, beta_e, accuracy, and off-task loss.
  3. For Qwen2.5-3B generated accuracy is conditional on parseable outputs. What are the parse rates with and without the intervention? If the intervention changes parse rate, the conditional accuracy gains could be misleading.
  4. Since beta_e is a slope over score gap while most of the margin sits in the intercept, in what sense does preserving 90% of beta_e guarantee that "comparison evidence" is preserved? Would an intervention that only rescales the margin pass your constraint?
  5. Can you show one experiment closer to the motivating use case (e.g., pairwise judging of two model responses with ground-truth preference labels), even for one model?
  6. Please check overlap with recent mechanistic work on LLM-judge bias that appeared before the deadline (e.g., arXiv 2607.11871, July 2026); I did not read it in detail, but the related-work section should position against it if relevant.
Reviewer_38gPOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose a mechanism for detecting hidden position bias: whether an LLM's preference depends partly on which option appears first or second, even when swapping the options doesn't change its final answer. They then try to locate and adjust the internal components that carry this influence, and reduce position bias while preserving the model's ability to compare the actual evidence correctly.

Critical strengths

The paper asks a useful diagnostic question: whether the order of the options can influence a model even when it keeps choosing the right answer? It then tests if this hidden positional influence can be reduced without hurting the model's ability to perform the underlying comparison.

I also liked that the authors make targeted changes inside model internals to test which parts affect the bias, which gives the analysis more weight than simply observing changes in the answers.

Critical weaknesses

Claim soundness and generality

I found the paper difficult to follow, and I didn't immediately see why the question was useful or what the main contribution was beyond prior work. In particular, I'd like a clearer explanation of why reducing these hidden order effects matters when the model is already getting the answer correctly. The claims also feel quite broad given the fairly slim experimental backing and limited range of models tested. In a future revision, I'd like to see the authors sharpen the motivation and main contributions, expand the experiments to more realistic settings, and be clearer about how far their findings actually extend. As it stands, the gap between the claims and the evidence makes me less confident in the broader conclusions.

I also wanted a clearer motivation for why the authors chose to test only instruction-tuned models. It would help to compare them with the same models before instruction tuning, to see whether the positional bias was already there, and if the same interventions would reduce it. Although the choice to focus on instruction-tuned models is relevant to LLM-as-a-judge applications, comparisons with pre-instruction-tuning matched checkpoints under comparable prompts would strengthen the analysis. Otherwise, the authors should more explicitly limit the generality of their conclusions.

Missing experiments: pairwise comparisons of actual model responses, and agreement with human judgments

I suggest evaluating the method on pairwise comparisons of actual model responses, using human preference labels, and including responses that are closely matched in quality. Comparing against a baseline that combines judgments from both presentation orders would help establish if the internal intervention adds any additional value for practical evaluation tasks.

The experiments do demonstrate reductions in positional influence and accuracy gains on controlled comparison tasks, but I’m still not sure how useful it would be for judging real model responses. For models that already get the numerical comparisons near-perfectly, reducing the bias does not improve their accuracy (Table 3), so the added benefit is unclear.

Presentation clarity

The figures and tables captions do not really provide clarity on how to contextualize the presented results, and I didn’t find sufficient discussion in the main text prose. Perhaps the authors could add some highlights or callouts to the captions to direct visual attention to the main effects.

Questions for authors

For models whose predictions are already nearly perfectly accurate in both presentation orders, what additional reliability problem does the latent positional effect expose or solve?

Can the authors demonstrate that reducing this effect improves agreement with human judgments on realistic response comparisons, beyond what's achievable by simply evaluating both presentation orders?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.