Truth as a Trajectory: What Internal Representations Reveal About Large Language Model Reasoning (2026.acl-long)
Copied to clipboard
| Challenge: | Existing explainability methods for Large Language Models treat hidden states as static points in activation space, but they are saturated with polysemantic features. |
| Approach: | They propose a framework that shifts analysis from static activations to layer-wise geometric displacement. |
| Outcome: | The proposed framework outperforms existing explainability methods on commonsense reasoning, question answering, and toxicity detection benchmarks. |
Similar Papers
Trajectory Signatures of Deception in Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Existing interpretability methods exhibit structural limitations in the field of deception detection. |
| Approach: | They propose a method to capture layerwise activations at sparse "decision points" . they capture deception as a dynamic process, a trajectory through the model's hidden-state space . |
| Outcome: | The proposed classifier achieves comparable performance to PCA-reduced probing for binary sycophancy detection and shows preliminary utility for 4-way deception-type classification. |
LLM Reasoning as Trajectories: Step-Specific Representation Geometry and Correctness Signals (2026.acl-long)
Copied to clipboard
| Challenge: | Existing models generate tokens by updating high-dimensional representations and decoding from them at each timestep. |
| Approach: | They propose a framework that allows reasoning correction and length control based on derived ideal trajectories. |
| Outcome: | The proposed model can predict correctness and length control based on ideal trajectories. |
Tracing Logit Trajectories Across Layer Depth: Dataset-Level Explainability for Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Sentence-level explanations miss the bigger picture of how a black-box model behaves across data . a dataset-level analysis that traces the intermediate structure of decision formation is needed . |
| Approach: | They propose a method that aggregates logit updates into a reproducible dataset-level trajectory pattern. |
| Outcome: | The proposed model enables depth-wise explainability across 6 languages and 5 NLP tasks. |
Dissecting Failure Dynamics in Large Language Model Reasoning (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models achieve strong performance through extended inference-time deliberation, yet how their reasoning failures arise remains poorly understood. |
| Approach: | They propose a framework that probes and redirects critical transitions using uncertainty signals. |
| Outcome: | Empirical evaluations show that GUARD improves reasoning performance . GUard probes critical transitions and redirects them using uncertainty signals . |
Probing the Geometry of Truth: Consistency and Generalization of Truth Directions in LLMs Across Logical Transformations and Question Answering Tasks (2025.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are trained on vast corpora that contain substantial knowledge but their outputs often contain confidently stated inaccuracies. |
| Approach: | They propose to encode truthfulness as a distinct linear feature, termed the "truth direction", which can classify truthfulness reliably. |
| Outcome: | The proposed model can generalize to logical transformations, question-answering tasks, in-context learning, and external knowledge sources. |
How Context Shapes Truth: Geometric Transformations of Statement-level Truth Representations in LLMs (2026.acl-long)
Copied to clipboard
| Challenge: | Prior work shows that large language models encode whether a statement is true as a vector in residual stream activations. |
| Approach: | They study how truth vectors change when context is introduced in Large Language Models . they measure directional change between truth vector with and without context and relative magnitude of truth vector upon adding context. |
| Outcome: | The results show that large models distinguish relevant from irrelevant context mainly through directional change () |
Bridging Internal Consistency and External Alignment: A Causal and Dynamic Interpretability Framework for LLM Generation (2026.acl-long)
Copied to clipboard
| Challenge: | Existing interpretability methods focus on internal and external aspects of the model . existing explanations often focus on surface correlations or static dependencies . |
| Approach: | They propose a causal and dynamic interpretability framework for Large Language Models . they characterize backdoor-adjusted causal effects of generated prefix and prompt . |
| Outcome: | The proposed framework provides a unified causal view of internal consistency and external alignment in LLM generation dynamics. |
Towards Intrinsic Interpretability of Large Language Models: A Survey of Design Principles and Architectures (2026.acl-long)
Copied to clipboard
| Challenge: | Existing studies on explainable AI focus on post-hoc explanation methods that interpret trained models through external approximations. |
| Approach: | They propose to categorize existing approaches into five design paradigms: functional transparency, concept alignment, representational decomposability, explicit modularization, and latent sparsity induction. |
| Outcome: | The proposed approaches are categorized into five design paradigms: functional transparency, concept alignment, representational decomposability, explicit modularization, and latent sparsity induction. |
FOL-Traces: Verified First-Order Logic Reasoning Traces at Scale (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing approaches to evaluate language models fail to provide structural clarity and verifiable inference. |
| Approach: | They propose to use a large-scale dataset of programmatically verified reasoning traces to evaluate structured logical inference. |
| Outcome: | The proposed model achieves 45.7% accuracy on masked operation prediction and 27% on two-step completion. |
TRACE: Training and Inference-Time Interpretability Analysis for Language Models (2025.emnlp-demos)
Copied to clipboard
| Challenge: | Existing tools for interpretability analysis of transformer models are post hoc, rely on scalar metrics or require nontrivial integration effort. |
| Approach: | They propose a modular toolkit for training and inference-time interpretability analysis of transformer models. |
| Outcome: | Experiments with autoregressive transformers show that TRACE reveals developmental phenomena overlooked by traditional scalar metrics such as loss or accuracy. |