When Attribution Patching Lies: Diagnosis and a Second-Order Correction

• URL: http://arxiv.org/abs/2606.09899v1
• Date: Fri, 05 Jun 2026 05:20:32 GMT
• Title: When Attribution Patching Lies: Diagnosis and a Second-Order Correction
• Abstract summary: We show that the dominant error stems from the non-linearities in the downstream network rather than local curvature at the patched component.<n>This insight yields three practical tools: (i) a reliability score to detect untrustworthy estimates, (ii) error bounds quantifying potential attribution mis-specifications, and (iii) a Hessian-vector-product correction that eliminates the leading-order error with only one additional backward pass.
• Score: 9.338965648455238
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: A central goal of mechanistic interpretability is to identify which internal components causally drive a language model's behavior. Because these importance estimates serve as the evidence for identifying circuits, systematic errors can lead to the misidentification of the underlying mechanisms. While activation patching provides a gold-standard causal metric, its computational cost is prohibitive at scale. Practitioners instead rely on attribution patching, a gradient-based, first-order approximation whose reliability remains poorly understood. In this work, we characterize the source of this unreliability, demonstrating that the dominant error stems from the non-linearities in the downstream network rather than local curvature at the patched component. This insight yields three practical tools: (i) a reliability score to detect untrustworthy estimates, (ii) error bounds quantifying potential attribution mis-specifications, and (iii) a Hessian-vector-product (HVP) correction that eliminates the leading-order error with only one additional backward pass. In evaluations across five model families (124M-9B parameters) and both random-token and naturalistic (name-swap) perturbations, HVP is the only second-order correction feasible at larger scale, where standard baselines like Integrated Gradients become computationally prohibitive. In comparative experiments, a multi-step HVP variant matches or exceeds the accuracy of Integrated Gradients at significantly lower compute, outperforming prior second-order baselines. These improvements lead to higher-fidelity circuit recovery on standard benchmarks and support a Screen-Flag-Fix workflow that targets computational effort only toward the components flagged as unreliable.

Related papers

• Semantic Fibers and Cross-Gram Interference: A Calculus of Safety Drift in Overcomplete Representations [0.0]
  A deployed language model may refuse a harmful request in English yet comply with its faithful translation.<n>We formalize this phenomenon through an audited equivalence relation.<n>We show that, for a declared quotient, representation, metric, feature dictionary, scoring head, threshold, and contrast model, the resulting safety drift admits an exact linear-algebraic characterization.
  arXiv  Detail & Related papers  (2026-09-14T00:23:32Z)
• Beyond Polarization: The Generative Constraint of Chain-of-Thought in Pointwise Reranking [28.78994214839467]
  In pointwise document reranking, Chain-of-Thought models typically underperform direct scoring models.<n>Our empirical study first confirms that this gap is stable across scales up to 32B parameters, ruling out model and data capacity confounders.<n>We then apply stress tests utilizing reinforcement learning, fine-grained supervision, and architectural decoupling to explicitly repair these deviations.
  arXiv  Detail & Related papers  (2026-08-31T07:51:49Z)
• Conformal Changepoint Localization and Root Cause Analysis with Corrupted Observations [55.76952683833966]
  This paper proposes weighted CONCH (W-CONCH) and weighted CROC (W-CROC) to reduce confidence set size when data may be corrupted.<n>Experiments on image-based and real-world changepoint and root-cause benchmarks show that uncertainty-based weighting substantially reduces confidence set size.
  arXiv  Detail & Related papers  (2026-07-29T05:16:59Z)
• Tailoring Strictly Proper Scoring Rules for Downstream Tasks: An Application to Causal Inference [48.78219918881965]
  We propose a principled framework for deriving task-specific strictly proper scoring rules by matching the local curvature of the downstream error metric.<n>We apply this to the Average Treatment Effect (ATE) estimation, deriving a closed-form loss and its corresponding canonical probability mapping.
  arXiv  Detail & Related papers  (2026-06-02T08:41:25Z)
• Distributionally Robust Transfer Learning with Structurally Missing Covariates, with Application to Cross-National Cardiac Arrest Prediction [6.540707401117924]
  We propose DRUM (underlineDtributionally underlineRobust underlineUn) as a framework that transfers prediction models to target populations.<n>We show that DRUM yields better-supervised predictions and improved clinical classification performance across sites.
  arXiv  Detail & Related papers  (2026-05-22T20:53:04Z)
• Discovery of Hidden Miscalibration Regimes [52.452902154360565]
  Models may be systematically overconfident on some kinds of inputs and underconfident on others.<n>We define the corresponding miscalibration field and propose a diagnostic framework for estimating it.<n>Our approach learns a calibration-aware representation of the input space and estimates signed local miscalibration by kernel smoothing in the learned geometry.
  arXiv  Detail & Related papers  (2026-05-13T13:07:50Z)
• Architecture-agnostic Lipschitz-constant Bayesian header and its application to resolve semantically proximal classification errors with vision transformers [0.0]
  This work presents an architecture-agnostic Lipschitz-constant Bayesian header that can be integrated into feature extractors such as vision transformers.<n>We also propose a novel metric to capture uncertainty and confidence across misclassification rates, as well as an adaptive arithmetic-mean fusion scheme.<n>Although computational costs increase due to Monte Carlo sampling, the method offers plug-and-play compatibility with pre-trained backbones.
  arXiv  Detail & Related papers  (2026-05-07T09:18:06Z)
• Correct Answers from Sound Reasoning: Verifiable Process Supervision for Language Models [94.68358825189738]
  We propose a post-training framework for verifiable domains that jointly optimize prediction accuracy and reasoning quality.<n>We evaluate VPS on chess, a controlled testbed where reasoning steps can be deterministically verified against engine signals.<n>VPS preserves accuracy while significantly improving reasoning quality, reducing win-rate error by up to 30% and restoring consistency to near saturation.
  arXiv  Detail & Related papers  (2026-04-03T15:19:46Z)
• Overconfident Errors Need Stronger Correction: Asymmetric Confidence Penalties for Reinforcement Learning [17.384089089363382]
  We identify a root cause that existing methods overlook: the uniform penalization of errors.<n>Current approaches treat all incorrect rollouts within a group identically.<n>We propose the Asymmetric Confidence-aware Error Penalty (ACE)
  arXiv  Detail & Related papers  (2026-02-24T22:46:43Z)
• PRIME: A Process-Outcome Alignment Benchmark for Verifiable Reasoning in Mathematics and Engineering [71.15346406323827]
  We introduce PRIME, a benchmark for evaluating verifiers on Process-Outcome Alignment verification.<n>We find that current verifiers frequently fail to detect derivation flaws.<n>We propose a process-aware RLVR training paradigm utilizing verifiers selected via PRIME.
  arXiv  Detail & Related papers  (2026-02-12T04:45:01Z)
• Geometric Calibration and Neutral Zones for Uncertainty-Aware Multi-Class Classification [0.0]
  This work bridges information geometry and statistical learning, offering formal guarantees for uncertainty-aware classification in applications requiring rigorous validation.<n> Empirical validation on Adeno-Associated Virus classification demonstrates that the two-stage framework captures 72.5% of errors while deferring 34.5% of samples, reducing automated decision error rates from 16.8% to 6.9%.
  arXiv  Detail & Related papers  (2025-11-26T01:29:49Z)
• Reinforcement Learning with Verifiable yet Noisy Rewards under Imperfect Verifiers [90.50039419576807]
  Reinforcement Learning with Verifiable Rewards (RLVR) trains policies against automated verifiers to avoid costly human labeling.<n>To reduce vulnerability to verifier hacking, many RLVR systems collapse rewards to binary $0,1$ during training.<n>This choice carries a cost: it introduces textitfalse negatives (rejecting correct answers, FNs) and textitfalse positives (accepting incorrect ones, FPs)
  arXiv  Detail & Related papers  (2025-10-01T13:56:44Z)
• Revisiting Multivariate Time Series Forecasting with Missing Values [65.30332997607141]
  Missing values are common in real-world time series.<n>Current approaches have developed an imputation-then-prediction framework that uses imputation modules to fill in missing values, followed by forecasting on the imputed data.<n>This framework overlooks a critical issue: there is no ground truth for the missing values, making the imputation process susceptible to errors that can degrade prediction accuracy.<n>We introduce Consistency-Regularized Information Bottleneck (CRIB), a novel framework built on the Information Bottleneck principle.
  arXiv  Detail & Related papers  (2025-09-27T20:57:48Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.