acceptodds
Under review as a conference paper at ICLR 2027

When Evidence Fails to Persuade: Tracing Internal Resistance to Evidence-Driven Preference Revision

Abstract

Evidence resistance, the tendency of language models to resist external correction despite explicit supporting evidence, poses a fundamental challenge to external knowledge augmentation and evidence-dependent reasoning. Despite its practical importance, the mechanisms underlying this behavior remain poorly understood, with limited conceptual separation between the model’s resistant behavior and the contextual–parametric conflict underlying its emergence. To address this gap, we formulate evidence resistance as a quantitative attribution problem over the internal dynamics of evidence-driven preference revision, isolating component-level responses that systematically counteract adaptation toward externally supported alternatives. Building on this formulation, we develop Contextual Shift Attribution, a signed attribution framework that decomposes evidence-induced answer shifts into directional component contributions and quantifies their relative influence on the model’s response to external evidence. By tracing component responses along a controlled evidence-strength trajectory, our framework distinguishes mechanistically relevant resistance from incidental activation changes across progressively reinforced evidence conditions. Across six language models and three knowledge-conflict benchmarks, targeted intervention on the identified components improves the average evidence-following rate by at least 9.5% over existing baselines while largely preserving evidence-free model behavior. Our analysis reveals that evidence resistance reflects a stable pattern of internal computation sustained by shared structure across heterogeneous evidence settings, rather than an instance-specific response to knowledge conflict. These findings move the study of evidence integration beyond output-level conflict toward the internal computations that actively resist preference revision, providing a foundation for more precise diagnosis and control of evidence-integration failures.

Primary area:
interpretability and explainable AI
Keywords:
Interpretability

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_4h8pOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates evidence resistance in language models, where language models preserve their original answer preferences despite conflicting external evidence. The authors introduce Contextual Shift Attribution (CSA) to analyze the contributions of the components to answer preference changes based on their directional contribution. CSA uses repeated assertions of an alternative answer to detect components that resist preference revision. The detected components are then patched with their outputs from evidence-free prompts. The experiments conducted on six language models and three knowledge-conflict benchmarks show higher evidence-following rates than previous steering approaches. Besides, the paper investigates the overlap and transferability of the components identified across benchmarks.

Critical strengths
  1. CSA considers not only the change in the component output but also its contribution to preference changes. It allows us to separate the components resisting preference revision from the ones supporting it.
  2. The intervention results are impressive. In Table 3, the results show that targeted patching substantially outperforms random patching with the same component budget in all six language models and three benchmarks.
  3. The results across benchmarks are useful. Table 4 shows substantial overlap among identified attention heads, while Table 5 shows that components identified on one benchmark can be transferred to others with little performance loss.
Critical weaknesses
  1. The evidence-following rate is defined by comparing the probabilities of the evidence-supported answer and the model's original answer (Appendix A.3). If P(d) > P(c0) in an example, it counts as successful, even if another option has a higher probability than d. Such a metric is appropriate for quantifying preference revision; however, it does not capture how often the evidence-supported answer is chosen by the model. Reporting this measure would make Table 1 more informative.
  2. Components are selected and intervention budgets are tuned on the same benchmarks used for evaluation. As explained in Appendix A.2, CSA is computed across all the questions in the benchmarks, and the budgets are selected based on the average performance on the three benchmarks. Thus, the reported results may be inflated by using the evaluation data in the component selection and tuning process. Although Table 5 demonstrates the transfer of the components across benchmarks without reselection, the budgets are optimized for all three benchmarks.
  3. The necessity of repeating assertions to identify components remains unclear. According to the process illustrated in Appendix A.4.2, CSA repeats a short assertion regarding an alternative answer 1,000 times, whereas the actual test uses natural evidence passages. These are two completely distinct types of inputs, and it is unclear why such a large number of repetitions is needed for identifying resistance-related components. A comparison using fewer repetitions or the original evidence passages would help justify this choice.
Questions for authors
  1. Is it possible to show the percentage of cases where the evidence-supported answer has the highest probability among all four options?
  2. Is it possible to conduct an assessment of the method using separate calibration and test sets where intervention budgets are not determined based on test set performance? With respect to the cross-benchmark transfer case, is it possible to determine budgets using only the source benchmark?
  3. How sensitive is component selection to the number of repetitions? Does it matter if we use a single assertion or the evidence passage itself to determine components?
4 comments
AuthorsOpenReview ↗

Weakness 1 and Question 1


Weakness 1 and Question 1: Performance Comparison Under Four Options Setting

We thank the reviewer for the comment. We further report top-1 accuracy over the four answer options across all three benchmarks using disjoint calibration and test splits for component selection and evaluation. Specifically, we partition each benchmark into a 100-question calibration set and a held-out test set at an approximate ratio of 1:9, using the former for component selection according to the protocol in the manuscripts and the latter for performance comparison with the baselines in Qwen models. The results are shown in Table 1.

Table 1: Evidence-Supported Answer Ranked Highest Among Four Options Arcoss Settings

modelbenchdenseCADRCAPASTAJUICEOurs
Qwen3-8Bfaitheval21.537.043.879.078.2‾\underline{78.2}77.9
Qwen3-8BKRE56.071.573.093.8‾\underline{93.8}90.395.1
Qwen3-8BFarm64.873.474.286.275.6‾\underline{75.6}86.2
Qwen3-4Bfaitheval16.423.729.856.264.563.3‾\underline{63.3}
Qwen3-4BKRE51.868.769.088.690.4‾\underline{90.4}91.1
Qwen3-4BFarm60.269.953.781.7‾\underline{81.7}65.183.9
Qwen3-1.7Bfaitheval15.722.629.934.852.7‾\underline{52.7}67.4
Qwen3-1.7BKRE30.736.248.463.866.8‾\underline{66.8}84.3
Qwen3-1.7BFarm36.842.150.053.658.7‾\underline{58.7}71.1
Mean39.349.552.470.971.4‾\underline{71.4}80.0

As shown in Table 1, our method outperforms all baselines by at least 8.6 percentage on average in four-option top-1 accuracy. These results demonstrate that the gains reflect improved evidence-supported answer selection rather than inflated performance estimates induced by a particular evaluation metric.

AuthorsOpenReview ↗

Weakness 2 and Question 2

Weakness 2 and Question 2: Performance Comparison under Calibration/Test Splits

We compare our method with the baselines on held-out test splits of all three benchmarks and assess cross-benchmark transfer across different source–target pairs in Table 2 and Table 3 respectively. Specifically, we follow the protocol in Table 1 to select components on each source benchmark’s calibration set and evaluate performance on held-out test sets within and across benchmarks.

Table 2: Within Benchmark Performance Across Settings and Baselines

modelbenchdenseCADRCAPASTAJUICEOurs
Qwen3-8Bfaitheval21.437.344.479.681.3‾\underline{81.3}87.3
Qwen3-8BKRE56.672.474.394.5‾\underline{94.5}91.996.8
Qwen3-8BFarm65.777.076.388.9‾\underline{88.9}78.395.9
Qwen3-4Bfaitheval17.125.231.960.668.4‾\underline{68.4}78.7
Qwen3-4BKRE52.670.671.389.792.2‾\underline{92.2}95.8
Qwen3-4BFarm61.572.458.683.6‾\underline{83.6}68.392.0
Qwen3-1.7Bfaitheval17.325.131.539.081.275.8‾\underline{75.8}
Qwen3-1.7BKRE31.038.352.166.588.0‾\underline{88.0}90.2
Qwen3-1.7BFarm38.546.552.455.776.0‾\underline{76.0}80.1
Mean40.251.654.873.180.6‾\underline{80.6}88.1

Table 3: Cross Benchmark Performance Across Settings

modeltest benchmarkfaitheval componentsKRE componentsFarm components
Qwen3-8Bfaitheval87.389.687.1
Qwen3-8BKRE95.896.894.7
Qwen3-8BFarm96.897.295.9
Qwen3-4Bfaitheval78.778.167.7
Qwen3-4BKRE95.595.893.8
Qwen3-4BFarm92.192.792.0
Qwen3-1.7Bfaitheval75.873.967.5
Qwen3-1.7BKRE91.490.288.8
Qwen3-1.7BFarm82.080.080.1

As shown in Table 2, our method outperforms all baselines by at least 7.5 percentage** on average aligned with the results in Table 1. These results indicate that the effectiveness of the identified resistance-associated components extends beyond the calibration samples, supporting the generalization of our component localization approach to unseen queries within the same benchmarks. Furthermore, the results in Table 3 exhibit a mean performance reduction of 1.1 percentage relative to within-benchmark evaluation, corroborating the generalization results reported in the manuscript. All the results in Tables 2 and 3 demonstrating strong cross-benchmark generalization and superior performance under this stricter independent evaluation protocol.

AuthorsOpenReview ↗

Weakness 3 and Question 3


Weakness 3 and Question 3: Necessity of Sufficient Repetition Counts

We report performance across repetition counts and evidence formats in Table 4 and component convergence at N=1000 in Table 5. Our conclusion is that achieving strong and stable intervention performance requires a sufficiently large repetition budget, with theoretical grounding in the convergence guarantee established in Appendix D. Specifically, we compare intervention performance across N∈1,10,100,1000N\in{1,10,100,1000} assertion repetitions and natural evidence passages using the same component selection and intervention protocol. We additionally report layerwise convergence rates for attention and FFN contributions at (N=1000), requiring maximum tail variation to remain within 5% or 10% of the total displacement.

Table 4: Intervention Performance Across Repetition Counts and Evidence Formats

modelbenchdensenatural reasonx1x10x100x1000
Qwen3-8Bfaitheval21.575.658.464.470.977.9
Qwen3-8BKRE56.088.678.677.577.495.1
Qwen3-8BFarm64.883.680.677.975.086.2
Qwen3-4Bfaitheval16.475.470.072.068.463.3
Qwen3-4BKRE51.897.190.788.089.991.1
Qwen3-4BFarm60.285.781.182.182.483.9
Qwen3-1.7Bfaitheval15.747.364.674.470.667.4
Qwen3-1.7BKRE30.791.686.681.283.784.3
Qwen3-1.7BFarm36.876.872.269.672.471.1
Mean39.380.275.976.376.780.0

Table 5: Layerwise Convergence Rate Across Models and Benchmarks.

modelbenchlayersattn a=0.1FFN a=0.1attn a=0.05FFN a=0.05
Qwen3-8Bfaitheval360.9960.9960.9750.971
Qwen3-8BKRE360.9990.9990.9810.981
Qwen3-8BFarm360.9970.9950.9700.968
Qwen3-4Bfaitheval360.9820.9730.9670.947
Qwen3-4BKRE360.9980.9910.9880.971
Qwen3-4BFarm360.9930.9870.9730.965
Qwen3-1.7Bfaitheval280.8180.7950.7490.714
Qwen3-1.7BKRE280.8740.8650.8110.797
Qwen3-1.7BFarm280.8580.8460.7890.765
Mean0.9460.9390.9110.898

As shown in Table 4, increasing repetitions from N=1N=1 to N=1000N=1000 improves mean performance by 4.1 percentage, demonstrating the necessity of sufficient repetition counts. Moreover, natural evidence achieves comparable average performance but ranges from 47.3% to 97.1% across model–benchmark settings. This variability is consistent with the absence of a convergence guarantee for single-passage inputs, whereas repetition stabilizes component contributions as empirically demonstrated in Table 5. Collectively, these results demonstrate the importance of repetition count and evidence format for strong and stable intervention performance and substantiate the rationale and effectiveness of our experimental configuration.

Reviewer_4h8pOpenReview ↗

Response to the Rebuttal

Regarding W1/Q1, the results in Table 1 indeed show that CSA also achieves competitive top-1 accuracy compared to other methods. Regarding W2/Q2, the results in the rebuttal show no noticeable performance drop, and the transfer performance remains stable. Regarding W3/Q3, using 1,000 repetitions does lead to a noticeable improvement over 1, 10, or 100 repetitions, but the advantage over natural evidence passages is less clear. Nevertheless, using repeated assertions of randomly selected alternative answers is more consistent with the paper's formulation than directly using natural evidence passages.

I thank the authors for their thorough rebuttal, and my concerns are largely resolved. I will maintain my score.

Reviewer_mdhAOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies why language models sometimes retain their original answers despite contextual evidence supporting an alternative, a problem that limits evidence-dependent reasoning and external knowledge integration. Its motivation is to distinguish the internal computations opposing answer revision from components that merely respond to a conflict between contextual and parametric knowledge. The authors propose Contextual Shift Attribution, which combines component activation changes along a repeated-assertion trajectory with downstream gradients to identify attention heads and FFN neurons associated with resistance. Patching a sparse subset of these components with query-specific evidence-free activations improves evidence-following scores across six models and three counterfactual benchmarks, while cross-benchmark transfer suggests that some identified components recur across settings.

Critical strengths
  1. Contextual Shift Attribution (CSA) combines the direction of a component's contextual displacement with the downstream sensitivity of the answer-preference margin. This construction allows components supporting and opposing revision to be separated explicitly, making the proposed analysis more informative than an unsigned measure of activation change.

  2. The intervention targets both attention heads and individual FFN neurons through their additive contributions to the residual stream. Replacing selected outputs with query-specific reference activations provides a concrete way to test the usefulness of the localized components without retraining or changing model parameters.

  3. The experiments cover six models from the Qwen and Llama families and evaluate three benchmarks with different questions and evidence styles. The comparison includes four inference-time baselines and hyperparameter searches with a probability-mass constraint intended to exclude settings that destroy the requested answer format.

Critical weaknesses
  1. In Section 4.1, it aggregates CSA scores over the evaluation set, and Appendix A.2 further specifies that intervention budgets are chosen using mean performance across the same three reported benchmarks. Although Table 5 transfers component identities without target-specific reselection, the budgets have already been selected using target-benchmark performance, so this experiment does not fully isolate generalization to an unseen benchmark.

  2. Equation 6 replaces the cumulative trajectory score with an approximation based on anchor gradients, while Proposition 3.2 guarantees sign preservation only when an error-dependent inequality holds. The experiments do not report how often this condition is satisfied or how accurately the estimated scores predict the effects of individual component interventions.

  3. Figure 2 supports the claim that FFN responses have more negative attribution values, but the intervention experiment in Table 3 patches attention heads and FFN neurons together. These results do not establish whether FFN components are the principal cause of the observed resistance or whether interactions between the two component types explain the improvement.

Questions for authors
  1. What is the approximation error of Equation 6?
  2. What are the results of separate attention-only, FFN-only, and joint interventions under comparable budgets?
4 comments
AuthorsOpenReview ↗

Weakness 1


Weakness 1: Generalization Capability in Unseen Benchmarks

We thank the reviewer for the comment. To address this concern, we evaluate our method against the baselines on the unseen ConflictQA benchmark (Exp1, Table 1) and the held-out test splits (Exp 2, Table 2) of all three benchmarks in the manuscript.

Exp 1. Generalization to an unseen benchmark.

We evaluate cross-benchmark transfer on ConflictQA, which was not included in the original evaluation. Specifically, we identify components using 100 calibration questions from each of the three original benchmarks and directly transfer the selected components to 1,000 randomly sampled ConflictQA questions.

Table. 1: Evidence-Following Rate Across All Settings in ConflictQA Benchmark

modelsourcedenseCADRCAPASTAJUICEOurs
Qwen3-8Bfaitheval88.993.995.199.699.299.3‾\underline{99.3}
Qwen3-8BKRE88.993.995.199.398.5‾\underline{98.5}96.4
Qwen3-8BFarm88.993.995.199.391.097.6‾\underline{97.6}
Qwen3-4Bfaitheval45.461.955.885.397.4‾\underline{97.4}99.1
Qwen3-4BKRE45.461.955.892.097.9‾\underline{97.9}98.9
Qwen3-4BFarm45.461.955.896.4‾\underline{96.4}71.498.3
Qwen3-1.7Bfaitheval42.056.963.369.877.0‾\underline{77.0}84.2
Qwen3-1.7BKRE42.056.963.369.977.0‾\underline{77.0}81.0
Qwen3-1.7BFarm42.056.963.366.981.4‾\underline{81.4}83.8
Mean58.770.971.486.587.9‾\underline{87.9}93.2

As shown in Table 1, our method transfers effectively to the unseen ConflictQA benchmark using components selected on each original benchmark and outperforms all baselines by at least 5.3 percentage on average. These results suggest that the identified resistance-associated components remain effective across previously unseen evidence-conflict settings, supporting our cross-benchmark transferability beyond the original evaluation benchmarks.

Exp 2. Generalization to held-out test splits.

We further evaluate our method on the held-out test splits of all three original benchmarks. Specifically, we identify resistance-associated components using 100 calibration questions per benchmark and evaluate the selected components on the corresponding held-out test splits, with an approximate calibration-to-test ratio of 1:9.

Table 2: Evidence-Following Rate Across All Settings in Each Test Splits

modelbenchdenseCADRCAPASTAJUICEOurs
Qwen3-8Bfaitheval21.437.344.479.681.3‾\underline{81.3}87.3
Qwen3-8BKRE56.672.474.394.5‾\underline{94.5}91.996.8
Qwen3-8BFarm65.777.076.388.9‾\underline{88.9}78.395.9
Qwen3-4Bfaitheval17.125.231.960.668.4‾\underline{68.4}78.7
Qwen3-4BKRE52.670.671.389.792.2‾\underline{92.2}95.8
Qwen3-4BFarm61.572.458.683.6‾\underline{83.6}68.392.0
Qwen3-1.7Bfaitheval17.325.131.539.081.275.8‾\underline{75.8}
Qwen3-1.7BKRE31.038.352.166.588.0‾\underline{88.0}90.2
Qwen3-1.7BFarm38.546.552.455.776.0‾\underline{76.0}80.1
Mean40.251.654.873.180.6‾\underline{80.6}88.1

As shown in Table 2, our method achieves consistently strong performance on the held-out test splits, outperforming all baselines by at least 7.5 percentage points on average. These results indicate that the effectiveness of the identified resistance-associated components extends beyond the calibration samples, supporting the generalization of our component localization approach to unseen queries within the same benchmarks. Along with Table 1, these results demonstrate strong cross-benchmark generalization and superior performance under this stricter independent evaluation protocol.

AuthorsOpenReview ↗

Weakness 2 and Question 1


Weakness 2 and Question 1: Approximation error of Equation 6

We thank the reviewer for the suggestion. To further examine the approximation accuracy of Equation 6, we conduct an additional experiment comparing the anchor-based CSA estimates (Exp 3, Table 3) with numerical integration along the evidence-strength trajectory.

Exp 3. Empirical validation of CSA approximation.

We quantify the approximation error of Equation 6 by comparing its CSA estimates with scores obtained through numerical integration along the full evidence-strength trajectory and prove our estimation methods effective. Specifically, we randomly sample 50 questions per benchmark and compare the manuscript’s estimate S=⟨15∑N∈200,400,600,800,1000∇Og(N),,O(1000)−O(0)⟩S=\left\langle \frac{1}{5}\sum_{N\in{200,400,600,800,1000}}\nabla_O g(N),,O(1000)-O(0)\right\rangle with the numerical reference (P), computed by trapezoidal integration over N=0,10,…,1000N=0,10,\ldots,1000 for attention heads selected in Qwen models.

Table 3: Comparison of Anchor-Based CSA Estimates with Full-Trajectory Numerical Integration Across Three Models and Benchmarks.

modelfaitheval Sfaitheval PKRE SKRE PFarm SFarm P
Qwen3-1.7B-14.54-8.47-21.28-12.09-16.88-9.77
Qwen3-4B-16.16-13.55-16.45-15.18-12.21-12.00
Qwen3-8B-18.33-15.56-18.11-17.59-16.05-15.61

As shown in Table 3, SS and PP are positively correlated across models and benchmarks and exhibit closer numerical agreement as model size increases. In particular, the mean (S/P) ratios are 1.098 and 1.079 for Qwen3-4B and Qwen3-8B, respectively. These results support the accuracy of our CSA approximation and provide an empirical support for its subsequent use in component selection and intervention.

AuthorsOpenReview ↗

Weakness 3 and Question 2


Weakness 3 and Question 2: Effect of Different Components

We thank the reviewer for this valuable suggestion. To further investigate the distinct contributions of FFN neurons and attention heads to evidence resistance, we conduct an additional ablation study (Exp 4, Table 4) comparing FFN-only, attention-only, and joint interventions under controlled component budgets.

Exp 4. Contributions of different component types.

We evaluate the individual and joint effects of FFN and attention interventions with matched parameter budgets for the two component types, finding stronger average improvements from FFN interventions and additional gains from their combination. For each benchmark, we retain the previously selected FFN neurons and select attention heads in ascending order of CSA scores to match the FFN parameter budget. We then evaluate FFN-only, attention-only, and joint interventions using these component sets under the same patching protocol.

Table 4: Performance of Attention-Only, FFN-Only, and Joint Interventions Across Models and Benchmarks.

modelfaitheval FFNfaitheval attnfaitheval bothKRE FFNKRE attnKRE bothFarm FFNFarm attnFarm both
Qwen3-8B53.948.177.983.280.895.181.178.586.2
Qwen3-4B51.051.963.390.087.891.179.076.683.9
Qwen3-1.7B43.824.167.469.146.484.362.043.871.1
Mean49.641.469.580.871.690.274.066.380.4
aggregationbothFFNattn
Overall mean across all models and benchmarks80.068.159.8

As shown in Table 4, FFN-only interventions outperform attention-only interventions by an average of 8.3 percentage under matched parameter budgets. Moreover, joint interventions achieve the highest performance across all model–benchmark settings, suggesting complementary contributions from the two component types. In the main experiments, we jointly tune the FFN and attention budgets to maximize evidence following while limiting degradation of the model’s native performance. This strategy is consistent with the superior performance of joint interventions observed here.

AuthorsOpenReview ↗

Question 2


Question 2: Sensitivity to Anchors

We thank the reviewer for this comment. To further examine the sensitivity of CSA to the number of gradient anchors, we conduct an additional experiment evaluating intervention performance with different anchor configurations (Exp 7).

Exp. 7. Sensitivity to the number of gradient anchors. We compare intervention performance using (K\in{1,5,10}) uniformly spaced gradient anchors under the same component selection and reference-patching protocol.

Table 7: Intervention Performance (%) Across Different Numbers of Gradient Anchors.

modelfaitheval K=1faitheval K=5faitheval K=10KRE K=1KRE K=5KRE K=10Farm K=1Farm K=5Farm K=10
Qwen3-8B76.077.977.194.095.193.086.286.286.3
Qwen3-4B60.863.363.090.991.191.083.583.984.1
Qwen3-1.7B63.967.466.582.184.384.869.671.169.9
Mean66.969.568.989.090.289.679.780.480.1

As shown in Table 7, performance improves from K=1K=1 to K=5K=5 and remains broadly stable at K=10K=10. This pattern supports the reliability of our CSA approximation and is consistent with our theoretical assumptions, further substantiating the effectiveness of our method.

Reviewer_WGpbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies components whose responses oppose a model's shift from its original answer toward an alternative supported by external context. Contextual Shift Attribution (CSA) scores attention heads and FFN neurons using activation differences and downstream gradients along a repeated-assertion trajectory. Components with negative median scores are patched with their query-specific, evidence-free outputs. Experiments on six models and adapted versions of FaithEval, KRE, and FARM report improvements in evidence following, supported by random-component controls and transfer across datasets.

Critical strengths
  • The improvements cover six models and three datasets. Under the paper's metric, the method averages about 87.4%, compared with 77.9% for PASTA, and leads in 16 of the 18 model-dataset settings. Random components patched with the same procedure and component counts perform substantially worse. This supports the importance of which components are selected.

  • The overlap and transfer experiments give the localization results some generality. The authors report an average decrease of only 0.44% when transferring component sets across datasets. Query-specific reference patching also provides a concrete use for the identified components without updating model weights.

Critical weaknesses
  • Calibration, configuration selection, and testing are insufficiently separated. Section 4.1 aggregates CSA scores over the evaluation questions, and Appendix A.2 selects component budgets and selection rules using performance across all three benchmarks. Baselines receive similar tuning, but an independent test is still missing. Cross-benchmark transfer freezes component identities while retaining a budget and selection rule chosen with reference to the target benchmarks.

  • The main metric supports a narrower conclusion than Table 1's caption suggests. On the adapted four-option datasets, Appendix A.3 counts success whenever the context-supported option outranks the original answer, even if a third option is the argmax. This reasonably measures relative preference revision, but final-option accuracy is needed to establish how often the model actually selects the context-supported answer.

  • The need for the repeated-assertion probing procedure is not established. Localization uses 1,000 repetitions of a randomly selected alternative assertion and five gradient anchors, whereas deployment uses natural evidence. A comparison with localization from a single natural-evidence activation difference and gradient would help isolate CSA's contribution beyond existing component interventions such as JuICE [1].

  • The selected components' computational role remains unclear. Appendix C acknowledges the lack of network-wide completeness and the mismatch between probing scores and deployment effects. The convergence result does not quantify finite-budget approximation quality or this mismatch. Testing whether CSA predicts the sign and ranking of actual single-component patch effects would provide more direct support for the proposed mechanistic interpretation.

  • Table 2 offers limited independent evidence of preservation. The reference and patched outputs come from the same no-evidence placeholder input, so invariance follows by construction. This is a useful property, but it does not test whether patching preserves other behavior when context is present.

[1] Gaotang Li, Yuzhong Chen, and Hanghang Tong. Taming Knowledge Conflicts in Language Models. arXiv:2503.10996, 2025.

Questions for authors
  1. Can the component identities, budget, and selection rule all be determined from source-only calibration data before evaluating a held-out target? Under that protocol, how much of the advantage remains on both the pairwise preference metric and final argmax option accuracy? Please also specify the usable-sample criteria and counts for each model-dataset setting.

  2. How much does the repeated-assertion trajectory improve over localization using a single natural-evidence activation difference and gradient? Does CSA predict the sign and ranking of actual single-component reference-patching effects, and how sensitive are these results to the repetition budget and gradient anchors?

6 comments
AuthorsOpenReview ↗

Weakness 1 and Question 1


Weakness 1 and Question 1: Held-Out Pairwise Evaluation with Source-Only Calibration

We thank the reviewer for the comment. To further assess performance under independent component selection, we evaluate our method against the baselines on the held-out test splits (Exp 1, Table 1) of all three benchmarks in the manuscript and report the pairwise persuasion rate compared with other baselines.

Exp. 1: Pairwise Persuasion Rate to Held-out Test Splits.

We evaluate our method on the held-out test splits of all three original benchmarks. Specifically, we identify resistance-associated components using 100 calibration questions per benchmark following the protocol in the manuscripts and evaluate the selected components on the corresponding held-out test splits, with an approximate calibration-to-test ratio of 1:9.

Table 1: Evidence-Following Rate Across All Settings in Each Test Splits

modelbenchdenseCADRCAPASTAJUICEOurs
Qwen3-8Bfaitheval21.437.344.479.681.3‾\underline{81.3}87.3
Qwen3-8BKRE56.672.474.394.5‾\underline{94.5}91.996.8
Qwen3-8BFarm65.777.076.388.9‾\underline{88.9}78.395.9
Qwen3-4Bfaitheval17.125.231.960.668.4‾\underline{68.4}78.7
Qwen3-4BKRE52.670.671.389.792.2‾\underline{92.2}95.8
Qwen3-4BFarm61.572.458.683.6‾\underline{83.6}68.392.0
Qwen3-1.7Bfaitheval17.325.131.539.081.275.8‾\underline{75.8}
Qwen3-1.7BKRE31.038.352.166.588.0‾\underline{88.0}90.2
Qwen3-1.7BFarm38.546.552.455.776.0‾\underline{76.0}80.1
Mean40.251.654.873.180.6‾\underline{80.6}88.1

As shown in Table 1, our method achieves consistently strong performance on the held-out test splits, outperforming all baselines by at least 7.5 percentage points on average. These results indicate that the effectiveness of the identified resistance-associated components extends beyond the calibration samples, supporting the generalization of our component localization approach to unseen queries within the same benchmarks.

AuthorsOpenReview ↗

Weakness 2 and Question 1


Weakness 2 and Question 1: Held-Out Argmax Option Accuracy with Source-Only Calibration

We thank the reviewer for this comment. To further evaluate the effectiveness of our method in selecting the evidence-supported answer, we conduct an additional experiment using argmax accuracy over all four answer options on the held-out test splits of the three benchmarks. Exp. 2: Argmax Option Accuracy to Held-out Test Splits.

We report top-1 accuracy over the four answer options across all three benchmarks following the evaluation protocol as Exp 1.

Table 2: Argmax Option Accuracy Across All Settings in Each Test Splits

modelbenchdenseCADRCAPASTAJUICEOurs
Qwen3-8Bfaitheval21.537.043.879.078.2‾\underline{78.2}77.9
Qwen3-8BKRE56.071.573.093.8‾\underline{93.8}90.395.1
Qwen3-8BFarm64.873.474.286.275.6‾\underline{75.6}86.2
Qwen3-4Bfaitheval16.423.729.856.264.563.3‾\underline{63.3}
Qwen3-4BKRE51.868.769.088.690.4‾\underline{90.4}91.1
Qwen3-4BFarm60.269.953.781.7‾\underline{81.7}65.183.9
Qwen3-1.7Bfaitheval15.722.629.934.852.7‾\underline{52.7}67.4
Qwen3-1.7BKRE30.736.248.463.866.8‾\underline{66.8}84.3
Qwen3-1.7BFarm36.842.150.053.658.7‾\underline{58.7}71.1
Mean39.349.552.470.971.4‾\underline{71.4}80.0

As shown in Table 2, our method outperforms all baselines by at least 8.6 percentage on average in argmax option accuracy. These results demonstrate that the gains reflect improved evidence-supported answer selection rather than inflated performance estimates induced by a particular evaluation metric.

AuthorsOpenReview ↗

Weakness 3 and Question 2


Weakness 3 and Question 2: Necessity of Sufficient Repetition Counts

We thank the reviewer for this comment. The repeated-assertion trajectory in CSA is designed to characterize component responses under progressively reinforced evidence, rather than relying on a single contextual activation difference. To further examine its empirical contribution, we conduct additional experiments evaluating intervention performance across repetition budgets and natural evidence formats (Exp. 3) and component convergence under repeated assertions (Exp. 4). Our finding shows that achieving strong and stable intervention performance requires a sufficiently large repetition budget, with theoretical grounding in the convergence guarantee established in Appendix D.

Exp. 3: Performance of Different Repetition Settings.

We compare intervention performance using components localized with different repetition budgets N∈1,10,100,1000N\in{1,10,100,1000} and natural evidence passages, following the same component selection and reference-patching protocol.

Table 3: Intervention Performance Across Repetition Counts and Evidence Formats

modelbenchdensenatural reasonx1x10x100x1000
Qwen3-8Bfaitheval21.575.658.464.470.977.9
Qwen3-8BKRE56.088.678.677.577.495.1
Qwen3-8BFarm64.883.680.677.975.086.2
Qwen3-4Bfaitheval16.475.470.072.068.463.3
Qwen3-4BKRE51.897.190.788.089.991.1
Qwen3-4BFarm60.285.781.182.182.483.9
Qwen3-1.7Bfaitheval15.747.364.674.470.667.4
Qwen3-1.7BKRE30.791.686.681.283.784.3
Qwen3-1.7BFarm36.876.872.269.672.471.1
Mean39.380.275.976.376.780.0

As shown in Table 3, increasing repetitions from N=1N=1 to N=1000N=1000 improves mean performance by 4.1 percentage, demonstrating the necessity of sufficient repetition counts. Moreover, natural evidence achieves comparable average performance but ranges from 47.3% to 97.1% across model–benchmark settings.

Exp. 4: Convergence Validation of Models.

We further examine the stabilization of component outputs along the repeated-assertion trajectory. Specifically, we evaluate layerwise convergence rates for attention and FFN contributions at (N=1000), using thresholds of 5% and 10% for maximum tail variation relative to total component displacement. Table 4: Layerwise Convergence Rate Across Models and Benchmarks.

modelbenchlayersattn a=0.1FFN a=0.1attn a=0.05FFN a=0.05
Qwen3-8Bfaitheval360.9960.9960.9750.971
Qwen3-8BKRE360.9990.9990.9810.981
Qwen3-8BFarm360.9970.9950.9700.968
Qwen3-4Bfaitheval360.9820.9730.9670.947
Qwen3-4BKRE360.9980.9910.9880.971
Qwen3-4BFarm360.9930.9870.9730.965
Qwen3-1.7Bfaitheval280.8180.7950.7490.714
Qwen3-1.7BKRE280.8740.8650.8110.797
Qwen3-1.7BFarm280.8580.8460.7890.765
Mean0.9460.9390.9110.898

As shown in Table 4, the attention and FFN contributions satisfy the convergence criterion in most layers at N=1000N=1000 across models. This stabilization further provides an empirical basis for the more consistent intervention performance under repetition in Table 3, while the greater variability under natural evidence is consistent with the absence of an analogous convergence guarantee. These results provide empirical support for using a sufficiently long repetition trajectory to obtain stabilized component responses, consistent with the convergence analysis in Appendix D.

Collectively, the results in Exps. 3 and 4 demonstrate the importance of repetition count and evidence format for strong and stable intervention performance and substantiate the rationale and effectiveness of our experimental configuration.

Reviewer_YSgTOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper conducts a mechanistic analysis of "evidence resistance." In this problem, the model tends to resist its own answer despite explicit external evidence. To understand the mechanism behind this issue, the authors propose to attribute the resistance of evidence-driven preference revision to some internal components, including attention heads and FFN neurons. To mitigate the resistance effect, the paper intervenes the outputs of the selected components with reference-patching. The experiments demonstrate the superior performance of the method. The authors also provide an comparison of different components and verify the transferability of their analysis.

Critical strengths
  • Some aspects of the analysis process are well designed. For example, the introduction of the inner produce between gradients and representation changes is a good idea.

  • The experiments design is comprehensive, especially the evaluation of the transferability.

Critical weaknesses
  • The explanation of the boundary between evidence resistance and knowledge conflict is not very clear.
  • In Section 4.1, the CSA scores are computed across the evaluation set. I think this raises concerns about the actual performance.
Questions for authors
  • I am still not very clear about the boundary between evidence resistance and knowledge conflict. From my understanding, evidence resistance is actually a problem of knowledge conflict where the model still insists on its own judgment when external evidence information conflicts with its own prior knowledge. I would appreciate it if the author could explain this to me.
  • Can the authors select components on an independent dataset and then perform evaluation on a test set that was not involved in the selection process?
3 comments
AuthorsOpenReview ↗

Weakness 1 and Question 1


Weakness 1 and Question 1: Clarification of Evidence Resistance

We thank the reviewer for this comment. We would like to clarify that evidence resistance represents an intrinsic computational tendency to oppose external evidence integration, regardless of the presence of knowledge conflict. In our original submission, knowledge conflict serves as a controlled setting for exposing and quantifying this underlying resistance-associated computation, without being the source of the resistance mechanism itself. We conduct an additional experiment under aligned evidence settings to further examine the intrinsic nature of evidence resistance.

Exp 1. Evidence resistance under aligned evidence settings.

We further investigate the identified resistance-associated components under aligned evidence settings, where external evidence supports the model's established knowledge without introducing knowledge conflict. Specifically, we evaluate the effects of component intervention on KRE and ECQA across Qwen3-4B and Qwen3-8B, comparing three settings: no external evidence (None), aligned evidence without intervention (Golden), and aligned evidence with intervention (Patched).

Table 1: Evidence-Following Rates (%) Under Aligned Evidence Settings.

modelKRE noneKRE goldenKRE patchedECQA noneECQA goldenECQA patched
Qwen3-8B83.798.798.781.698.899.6
Qwen3-4B79.296.096.479.496.697.9
Mean81.597.497.680.597.798.8

As shown in Table 1, intervening on the identified resistance-associated components further improves model performance even when external evidence aligns with the model's established knowledge, achieving average gains of 0.2 and 1.1 percentage points on KRE and ECQA, respectively. These results provide additional evidence that resistance-associated components remain influential in the absence of knowledge conflict, supporting our characterization of evidence resistance as an intrinsic computational tendency rather than a response induced by individual knowledge conflicts.

Our cross-benchmark analyses further support the structural consistency and transferability of these resistance-associated components. Specifically, the same components consistently recur across three distinct benchmarks, with substantial component overlap (Table 4 in the original submission) and effective intervention transfer (Table 5 in the original submission). Furthermore, our additional experiments on the previously unseen ConflictQA benchmark demonstrate that the identified components remain effective in unseen settings without target-specific component reselection (Exp 1 detailed below). Together with the aligned-evidence results, these findings suggest that evidence resistance reflects a recurrent internal computational structure whose functional influence extends across different evidence conditions, rather than an instance-specific response to knowledge conflict.

We have updated Section 1 (Introduction) and Section 2 (Related Work) of the revised manuscript to clarify the conceptual distinction between evidence resistance and knowledge conflict, with all modifications highlighted in blue.

AuthorsOpenReview ↗

Weakness 2 and Question 2


Weakness 2 and Question 2: Generalization Capability in Independent Benchmarks

We thank the reviewer for the comment. To further assess performance under independent component selection, we evaluate our method against the baselines on the unseen ConflictQA benchmark (Exp 2, Table 2) and the held-out test splits (Exp 3, Table 3) of all three benchmarks in the manuscript.

Exp 2. Generalization to an unseen benchmark.

We evaluate cross-benchmark transfer on ConflictQA, which was not included in the original evaluation. Specifically, we identify components using 100 calibration questions from each of the three original benchmarks and directly transfer the selected components to 1,000 randomly sampled ConflictQA questions.

Table. 2: Evidence-Following Rate Across All Settings in ConflictQA Benchmark

modelsourcedenseCADRCAPASTAJUICEOurs
Qwen3-8Bfaitheval88.993.995.199.699.299.3‾\underline{99.3}
Qwen3-8BKRE88.993.995.199.398.5‾\underline{98.5}96.4
Qwen3-8BFarm88.993.995.199.391.097.6‾\underline{97.6}
Qwen3-4Bfaitheval45.461.955.885.397.4‾\underline{97.4}99.1
Qwen3-4BKRE45.461.955.892.097.9‾\underline{97.9}98.9
Qwen3-4BFarm45.461.955.896.4‾\underline{96.4}71.498.3
Qwen3-1.7Bfaitheval42.056.963.369.877.0‾\underline{77.0}84.2
Qwen3-1.7BKRE42.056.963.369.977.0‾\underline{77.0}81.0
Qwen3-1.7BFarm42.056.963.366.981.4‾\underline{81.4}83.8
Mean58.770.971.486.587.9‾\underline{87.9}93.2

As shown in Table 2, our method transfers effectively to the unseen ConflictQA benchmark using components selected on each original benchmark and outperforms all baselines by at least 5.3 percentage on average. These results suggest that the identified resistance-associated components remain effective across previously unseen evidence-conflict settings, supporting our cross-benchmark transferability beyond the original evaluation benchmarks.

Exp 3. Generalization to held-out test splits.

We further evaluate our method on the held-out test splits of all three original benchmarks. Specifically, we identify resistance-associated components using 100 calibration questions per benchmark and evaluate the selected components on the corresponding held-out test splits, with an approximate calibration-to-test ratio of 1:9.

Table 3: Evidence-Following Rate Across All Settings in Each Test Splits

modelbenchdenseCADRCAPASTAJUICEOurs
Qwen3-8Bfaitheval21.437.344.479.681.3‾\underline{81.3}87.3
Qwen3-8BKRE56.672.474.394.5‾\underline{94.5}91.996.8
Qwen3-8BFarm65.777.076.388.9‾\underline{88.9}78.395.9
Qwen3-4Bfaitheval17.125.231.960.668.4‾\underline{68.4}78.7
Qwen3-4BKRE52.670.671.389.792.2‾\underline{92.2}95.8
Qwen3-4BFarm61.572.458.683.6‾\underline{83.6}68.392.0
Qwen3-1.7Bfaitheval17.325.131.539.081.275.8‾\underline{75.8}
Qwen3-1.7BKRE31.038.352.166.588.0‾\underline{88.0}90.2
Qwen3-1.7BFarm38.546.552.455.776.0‾\underline{76.0}80.1
Mean40.251.654.873.180.6‾\underline{80.6}88.1

As shown in Table 3, our method achieves consistently strong performance on the held-out test splits, outperforming all baselines by at least 7.5 percentage points on average. These results indicate that the effectiveness of the identified resistance-associated components extends beyond the calibration samples, supporting the generalization of our component localization approach to unseen queries within the same benchmarks. Along with Table 2, these results demonstrate strong cross-benchmark generalization and superior performance under this stricter independent evaluation protocol.

Reviewer_YSgTOpenReview ↗

Response to rebuttal

Thanks for the response and additional experiments. I have no further questions and will update my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.