6arXiv cs.CL (Computation and Language)·
Causal audit of neuron selectors reveals refusal behavior lives in redundant subspace
A new arXiv preprint introduces a causal auditing framework for neuron-row attribution methods in LLMs, using one-shot zeroing interventions to test whether high-scoring neurons are genuinely causally important. The authors find that attribution methods outperform activation/magnitude baselines for identifying dispensable rows, and that refusal behavior (for hate/crime prompts) can be installed or removed via attributed rows while preserving fluency. A key finding is that refusal occupies a redundant subspace — different attribution methods identify largely disjoint sufficient row sets — meaning no single method recovers a unique causal mechanism. Rank-stability, a common proxy for selector quality, is shown to be a poor indicator of causal validity.
Related guides (3)
Related events (8)
Causal audit finds routing statistics do not predict expert importance in MoE pruning
A new arXiv paper conducts a token-level interventional audit of Mixture-of-Experts (MoE) pruning heuristics across three architectures (OLMoE-1B-7B, Qwen1.5-MoE, DeepSeek-V2-Lite), finding that no standard observational metric — utilization rates, activation norms, routing weight distributions — reliably predicts which experts can be removed without functional cost. Effect sizes fall below Cohen's d = 0.17 across all 60 metric-layer combinations after multiple-comparison correction, with only a single significant signal at OLMoE's final layer. The authors argue that existing pruning methods succeed not because they identify dispensable experts but because early-layer redundancy makes most selection criteria interchangeable. The work frames this as a concrete counterexample to the broader interpretability practice of treating associational (rung-1) evidence as interventional (rung-2) conclusions.
Mechanistic interpretability study reveals geometric structure of LLM-as-judge bias in hidden states
A new arXiv preprint investigates LLM-as-judge scoring bias at the representation level rather than the input-output level, studying seven judge models across seven bias types and nine benchmarks. The authors find that biased inputs are displaced along low-dimensional, type-specific subspaces in activation space, and that steering hidden states along these subspaces causally controls scoring direction. A linear projection onto bias-direction features predicts judge failures on unseen benchmarks, substantially outperforming text-based alternatives. The work provides a mechanistic account that unifies geometric structure, causal control, and operational prediction within a single framework.
Counterfactual Report Coordinates method for certifying incentive-compatible LLM honesty
A new arXiv paper introduces Counterfactual Report Coordinates (CRC), a method for identifying and controlling low-rank activation-space coordinates governing a model's answer, confidence, and caveat outputs, with the goal of making LLMs internally incentive-compatible — resistant to sycophantic pressure while remaining responsive to genuine evidence. The authors use causal interchange interventions (not probe accuracy) to identify these coordinates and introduce a training-free two-pass clamp that achieves near-perfect resist and update scores on a Bayesian-witness benchmark. Results transfer across three model families and to SycophancyEval, though the deployable single-pass compilation is lossy. The work frames activation-level counterfactual incentive-invariance as a structural primitive for alignment research.
CausaLab: Scalable Benchmark for Interactive Causal Discovery by LLM Agents
CausaLab is a new evaluation environment that tests LLM agents on interactive causal discovery tasks, requiring them to recover both causal graphs and structural equations from synthetic laboratory episodes governed by randomly sampled structural causal models (SCMs). The benchmark separates predictive accuracy from genuine causal understanding, revealing a persistent gap: GPT-5.2-high achieves 92% task accuracy in a 6-node observational setting but only 0.471 all-edge F1 for mechanism recovery. Mixed observation-intervention strategies improve structural fidelity, while pure intervention strategies underperform on both metrics. Premature stopping is identified as a key agent weakness, partially mitigated by prompting models to verify hypothesis-data consistency.
Alignment tuning installs distinct representational directions for sycophancy and cue-induced biases in LLMs
A new arXiv paper investigates where sycophancy and related cue-induced biases live inside LLMs, studying five model families and seven bias types. The authors find that alignment tuning — not pretraining — is primarily responsible for installing these biases, as base models show minimal susceptibility and no cue-specific activation signal. Each bias manifests as a distinct, causally active direction in hidden states that can be probed, transferred across datasets, and steered to recover unbiased answers. The work also demonstrates a modest debiasing intervention that reduces bias-induced errors while preserving correct answers.
MedFocus: Causal Visual Attribution Framework for Chest X-ray Reasoning in Large Vision-Language Models
This paper addresses the faithfulness of visual attribution methods in Large Vision-Language Models (LVLMs) applied to chest X-ray (CXR) analysis. The authors develop a causal evaluation framework using counterfactual editing to verify whether expert-annotated regions are causally responsible for model predictions, testing 11 attribution methods across six open-source LVLMs. Finding that existing attribution methods frequently fail to identify the actual visual evidence used by models, they propose MedFocus, a concept-based attribution method using unbalanced optimal transport to localize anatomical regions and measure their causal effect on outputs. MedFocus substantially outperforms prior methods and provides spatial, concept-level, and token-level attributions.
Analysis-driven transformer linearization outperforms prior baselines on LLaMA and Qwen up to 32B
A new arXiv paper analyzes why post-hoc linearization of causal self-attention degrades model quality, identifying key-dependent rank-1 orthogonal projections as the mechanism softmax relies on and explaining why delta-style networks outperform gated accumulation. The authors introduce structural interventions—sink tokens, short convolutions, and fixed-budget cache routing—applied in a frozen-backbone regime. Scaling across LLaMA and Qwen models up to 32B parameters, the approach outperforms prior post-hoc linearization baselines on MMLU and matches long-context retrieval of adaptive-caching frameworks.
OAT: Unsupervised failure attribution for LLM agents via one-class learning on successful trajectories
A new arXiv preprint introduces OAT, a method for attributing failures in LLM-based agentic systems by training exclusively on successful trajectories using one-class learning with neural controlled differential equations. At inference time, each step in a failure trajectory receives an anomaly score based on deviation from learned success dynamics. OAT is reported to be 200–5000× faster than prompting-based baselines while outperforming them by +20% F1 in-domain and +7% out-of-distribution, using only 100 successful trajectories for training. The approach addresses a practical bottleneck in agentic system debugging by eliminating the need for costly step-level failure annotations.