๐ฌ Interpretability¶
๐ฌ EMNLP2026 ยท 19 paper notes
๐ Same area in other venues: ๐ง NeurIPS2026 (106) ยท ๐๏ธ ECCV2026 (32) ยท ๐ท CVPR2026 (34) ยท ๐ฌ ICLR2026 (197) ยท ๐ฌ ACL2026 (63) ยท ๐งช ICML2026 (92)
๐ฅ Top topics: LLM ร3 ยท Reasoning ร2 ยท Alignment/RLHF ร2
- Attribute-Based Activation Steering of LLMs for Group-Specific Explanation Generation
-
This paper introduces an unsupervised, training-free attribute-based activation steering approach that extracts style and knowledge attributes of target groups from community Q&A texts and injects weighted activation vectors at inference time, achieving significantly improved group specificity while preserving factual accuracy.
- Beyond Decodability: Reconstructing Language Model Representations with an Encoding Probe
-
Addressing the core limitations of classic decoding probesโnamely non-comparable cross-feature metrics and vulnerability to feature collinearityโthis paper introduces an Encoding Probe that reverses the probing direction to reconstruct internal model activations using multivariate ridge regression, enabling direct feature contribution quantification on a shared unexplained variance (UV) scale.
- Confidence Before Answering: A Paradigm Shift for Efficient LLM Uncertainty Estimation
-
Pioneers a confidence-first paradigm and proposes CoCA, an end-to-end GRPO framework that dynamically grounds confidence targets in group-wise empirical success rates and decouples optimization via segmented credit assignment, reducing confidence inference token costs by over 92% while maintaining superior calibration.
- Dissecting Transformers: A CLEAR Perspective Towards Green AI
-
Addressing the severe temporal mismatch between microsecond-scale component execution and millisecond-scale hardware energy sensors, this work introduces CLEARโa three-stage profiling framework using activation caching and repetition amplificationโto conduct the first comprehensive component-level inference energy dissection across 15 models and 4 Transformer architectures.
- Diverging Transformer Predictions for Human Sentence Processing: A Comprehensive Analysis of Agreement Attraction Effects
-
Evaluating eleven autoregressive Transformers across comprehensive agreement attraction paradigms using surprisal theory reveals that while models align with human reading times in prepositional phrase modifiers, their predictions diverge sharply and contradict human cognitive asymmetries in object-extracted relative clauses, challenging the view that Transformers are faithful cognitive models of human morphosyntax.
- Do Compounds and Phrases Follow the Same Combinatorial Principles? What a Multi-Dimensional Exploration Reveals
-
Constructs COMPPHRASE-7, a parallel Chinese compound-phrase benchmark across seven construction types, and establishes a three-dimensional probing protocol disentangling feature transfer, category geometry, and causal specificity to demonstrate that Chinese PLMs encode partially transferable construction cues while category geometry is layer-sensitive and localized causal circuits are absent.
- Does Fine-Tuning Undo Activation Steering? Behavioural Recovery Without Weight-Edit Reversal
-
This paper investigates whether embedding activation steering directly into model weights survives downstream non-adversarial fine-tuning (SFT and RLHF), demonstrating that while fine-tuning can significantly restore steered behaviors (e.g., recovering 69% of ablated refusal under SFT), it reverses less than 0.8% of the weight edit, as the model routes around the intact edit via near-orthogonal weight updates.
- From Profiles to Steering Vectors: Global Sparse Priors and Local Semantic Calibration for Interpretable LLM Steering
-
Addressing semantic residue and context overfitting in retrieval- and activation-based personalization, this paper proposes GLASS, a training-free framework that extracts global style sparse priors via Sparse Autoencoders and constructs local contrastive vectors across scenario clusters, enabling lightweight, interpretable, dual-stream activation steering during inference.
- LSC-Pattern: Interpretable Lexical Semantic Change Analysis via Diachronic Sense Distributions
-
Addressing the black-box nature of conventional lexical semantic change detection that outputs only uninterpretable scalar change scores, LSC-Pattern connects LLM-based dictionary-grounded word sense disambiguation with multi-dimensional semantic properties (valence, semantic scope, metaphoricity), achieving high correlation with human judgments and transparent linguistic explanations across 150 years of Japanese corpora.
- Observable Patterns Are Not Explanations: A Causal-Geometric Analysis of Latent Reasoning Models
-
Challenging the widespread assumption that observable latent-state patterns (such as BFS tree search and scratchpad computation) constitute evidence for internal reasoning mechanisms, this paper uses matched controls, causal tracing, and gradient-subspace interventions to show that these patterns are epiphenomena rather than mechanisms, proving that latent-thought influence is graded and concentrated within low-dimensional geometric subspaces.
- SAEExplainer: Interpreting SAE Features with Activation-Guided Preference Optimization
-
Addressing the pervasive issues of over-generalization and explanation hallucinations caused by open-loop paradigms in Sparse Autoencoder (SAE) interpretation, SAEExplainer establishes a closed-loop framework that operationalizes objective model activation scores into verifiable rewards, achieving continuous self-refinement and superior causal triggering through two rounds of iterative Direct Preference Optimization (DPO).
- TBLS: Transition-Based Language Similarity through Activation Perturbation
-
To address the cross-model instability and static nature of conventional geometric language similarity measures, this paper proposes Transition-Based Language Similarity (TBLS), which injects controlled Gaussian noise into intermediate activations to trigger cross-lingual output transitions, benchmarking representation proximity via calibrated escape noise.
- The Mirage of Explainability: A Survey on Chain-of-Thought Faithfulness in Large Language Models
-
This paper delivers a comprehensive survey on the crisis of Chain-of-Thought (CoT) faithfulness in large language models, unifying its formal definition via dual internal alignment and external consistency criteria, establishing a four-stage failure taxonomy and four-level grounding benchmark hierarchy, and charting architectural decoupling pathways toward trustworthy AI.
- Universal Activation Verbalizer: A Unified Framework for Cross-Model Activation Explanation
-
Universal Activation Verbalizer (UAV) introduces a shared-decoder framework that leverages lightweight cross-attention adapters to map internal activations from heterogeneous donor models into soft tokens, enabling open-ended natural language explanation across model families and scales while slashing training GPU hours through adapter-only transfer.
- Weakening Neurons: An Input-Output Functionality in Transformers with Outsize Influence
-
By computing cosine similarities across input, gate, and output weight vectors of GLU neurons, this paper formalizes an input-output functionality taxonomy across modern LLMs, discovering that late layers contain a small cluster of highly active "weakening neurons" that unexpectedly sharpen next-token probability distributions via negative gate activations.
- What Drives Representation Steering? A Mechanistic Case Study on Steering Refusal
-
This paper introduces an activation patching framework tailored for multi-token autoregressive generation, revealing that refusal steering vectors derived from distinct methodologies rely on functionally interchangeable subnetwork circuits, demonstrating that steering primarily drives behavioral shifts through the attention OV circuit (specifically an input-invariant steering value vector) rather than the QK circuit, and achieving 85%โ96% dimension sparsification.
- Why and When to Use Statistical Significance in Explanation Evaluation: A Systematic Analysis
-
Systematically auditing 494 verified explainable NLP papers across the ACL Anthology (2020โ2025) reveals a persistent rigour gap where only 25.30% substantiate evaluation claims with statistical significance, motivating a risk-proportionate bipartite mapping framework between claims and testing regimes.
- Why Are Some Emotions Harder for LLMs? Uncovering the Causal Mechanisms of Emotion Inference via Sparse Autoencoders
-
By decoding monosemantic causal emotion features from residual streams using Sparse Autoencoders (SAEs), this paper uncovers why LLMs systematically struggle with disgust and surprise, providing a mechanistic explanation and introducing a lightweight sparse feature steering method that boosts weak emotions without degrading general capabilities.
- XQDT: eXplainable and Quantitative Data-Text Alignment Metric with Feedback Signals
-
XQDT introduces an end-to-end explainable and quantitative bidirectional data-text alignment metric that fine-tunes compact open-source models on semantically perturbed data to explicitly predict atomic unit labels (Correct, Omitted, Incorrect, Extra), aggregates these counts into closed-form precision, recall, and F1 scores, and provides structured feedback signals for iterative generation refinement.