The paper introduces Concept-Targeted Attribution (CTA), a method that trains attribution graphs to explain the emergence of internal concept representations in language models, rather than just the final token prediction. CTA produces probe-specific circuits that reveal which internal computations drive a linear probe’s accuracy, and cross-layer transcoders demonstrate that these graphs contain predictive structure across multiple concept categories. Causal ablations show that probe-targeted and logit-targeted graphs capture distinct mechanisms, with probe-relevant features affecting internal concept scores and logit-relevant features altering generated tokens.
By Vedant Palit, Florent Draye, Terry Jingchen Zhang, Bernhard Sch\"olkopf, Zhijing Jin
arXiv:2608. 10537v1 Announce Type: new Abstract: Sparse autoencoders (SAEs) have helped uncover mechanistic explanations for LLM behaviours such as reasoning, jailbreaking etc.
By Chuqiao Lin, Shivaji Sondhi, Xiao-Liang Qi
The paper introduces a method for identifying the dominant knowledge source behind large language model (LLM) outputs, distinguishing between faithfulness violations (misuse of provided context) and factuality violations (errors in internal knowledge). A simple linear probe trained on hidden representations can reliably classify this source, and the authors present AttriWiki, a self‑supervised pipeline that generates labeled training data by prompting models to recall withheld entities or read them from context. Probes trained on AttriWiki achieve high Macro‑F1 scores across several models and datasets, generalize zero‑shot to a benchmark, and show that attribution mismatches can increase error rates by up to 70%.
"whyItMatters":"The study demonstrates that knowing the source of an LLM’s answer is crucial for effective mitigation of hallucinations, as attribution mismatches significantly raise error rates."
By Ivo Brink, Alexander Boer, Dennis Ulmer
The paper investigates automated failure attribution (AFA) in large language model (LLM)-based agentic systems, where multiple agents collaborate on complex tasks. It finds that LLMs contain sufficient internal knowledge to attribute failures but often fail to output the correct step, revealing a knows‑but‑won’t‑say gap. To address this, the authors introduce grafting—a manual intervention that leverages latent model knowledge—showing it consistently outperforms direct prompting across several LLMs and benchmarks.
By Wendy Zheng, Liang Wu, Yinhan He, Mayank Darbari, Jundong Li, Liangjie Hong
arXiv:2607. 22610v1 Announce Type: new Abstract: When a language model produces a response in a multi-turn conversation, which tokens from prior turns shaped that answer, and how did those dependencies propagate across prior turns?
By Jessica Tang, Shraddha Barke, Sharad Agarwal
arXiv:2609.33676v2 Announce Type: replace
Abstract: LLM agents increasingly take consequential actions through interactions with users, policies, and external tools. Auditing these agents requires au...
By Yifan Liu, Praveen Venkateswaran, Abdulhamid Adebayo, Dong Wang
MEA is a multi‑agent framework that automates the selection and configuration of explanation tools for machine learning models across tabular, text, and vision data. A Proposer agent chooses appropriate tools based on the question and modality, while an Actor agent is trained end‑to‑end for faithfulness, producing natural‑language explanations. Experiments show that MEA outperforms existing post‑hoc explainers and closed‑source baselines, achieving significant faithfulness gains on six datasets.
By Yuyang Cheng, Raghav Kaushik Ravi, Srivarshinee Sridhar, Sriparna Saha, Akash Ghosh, Chirag Agarwal
arXiv:2607. 19361v1 Announce Type: cross Abstract: Most safety guardrails for large language models (LLMs) evaluate each prompt-response pair in isolation, which misses failures that arise only over a dialogue as benign turns compose into harm.
By Sanjay Mishra, Divya Chukkapalli, Ganesh R. Naik
arXiv:2606. 29171v1 Announce Type: cross Abstract: While existing data attribution methods can identify which training examples build specific mechanistic circuits, they cannot explain how training data shapes the high-level behavioral decisions a model learns to make.
By Reza Habibi, Darian Lee, Magy Seif El-Nasr
arXiv:2608. 09209v1 Announce Type: cross Abstract: Neural language models trained on large crowdsourced corpora frequently exploit spurious surface patterns tied to target labels without true linguistic or causal relevance, boosting benchmark performance while failing on adversarial or out-of-distribution inputs.
By Chidaksh Ravuru, Shashank Srivastava
The paper introduces SEAV, a verification‑centric framework for evaluating jailbreak attempts against large language models. SEAV decomposes responses into ordered steps and checks both validity and correctness using LLM‑as‑a‑judge and retrieval‑grounded verification. The method reduces false positives by 14.9 percentage points on a strategic‑dishonesty diagnostic and reclassifies 22.1–51.0% of previously successful jailbreaks as invalid across multiple benchmarks.
By Qilong Wu, Sahil Wadhwa, Pranab Mohanty, Giri Iyengar, Varun Chandrasekaran
The paper introduces the Controlled Attribution Benchmark (CATT), a synthetic evaluation framework that fine‑tunes language models on entirely fictitious knowledge to provide clear ground truth for data attribution. By systematically varying single dataset properties—such as document redundancy and lexical distractors—CATT reveals that these factors consistently degrade attribution performance. The study also explores two underexamined design choices: using negative log‑odds gradients instead of cross‑entropy gradients, and aggregating only the top token attribution scores, which together improve ground‑truth recall by 35–72 percentage points and enhance counterfactual retraining performance on FTRACE‑TREX by up to 4.9×.