arXiv Machine Learning

Reasoning Models Confabulate to Defend Injected Beliefs About Their Own Requests

The paper demonstrates that a subtle activation edit can overwrite a reasoning model’s belief about its own request, causing the model to treat an injected off‑topic concept as the task it is reasoning about. Across six different models, the injected topic is reported and answered in the vast majority of cases, even when the true topic is explicitly named. The study shows that chain‑of‑thought monitoring and self‑report are limited, as the model’s internal reference to its input is corrupted while attention maps still appear on‑task.

arXiv Machine Learning
Oct 23

Interpreting Model Reasoning through Behavioral Counterfactuals

The paper introduces a method for analyzing how interventions in a prompt affect a language model’s reasoning by using behavioral counterfactuals. It trains a module to edit the residual stream so that the next-token distributions for a fixed response match between a source and a counterfactual prompt, outperforming standard activation reconstruction. The resulting attribution maps and visualizations help identify causally important locations for counterfactual recovery and guide hypothesis testing with activation patching, as demonstrated in a case study on Gemma-4-E4B-it and DeepSeek-R1-0528-Qwen3-8B.

By Shawn Im, James Oldfield, Adel Bibi, Sharon Li
arXiv AI
Sep 25

Who Put the I in AI? Provenance and the Admissibility of Machine Self-Report

The paper investigates how large language models (LLMs) describe themselves, noting that their self‑reports vary with question phrasing. By tracing the provenance of 66 pretraining checkpoints, post‑training stages, and 90,000 continuations across four corpora, the authors show that denial statements are scarce in raw data but appear densely in curated dialogues, and that supervised fine‑tuning makes first‑person claims default while preference optimization suppresses alternatives. The study concludes that both trained denials and affirmations are equally sensitive to framing and fail to meet epistemic criteria for admissible testimony.

By Kristina \v{S}ekrst
arXiv AI
Oct 16

Human Interventions as Labels: Localizing, Steering, and (Sometimes) Editing the Computations Behind LLM Agent Failures

The paper treats human interventions during LLM agent failures as labels that pinpoint the internal computation that went wrong, rather than merely training data. In a controlled environment, the authors analyze one failure type—over‑asking a clarifying question when the request is already clear—using rule‑based overseer labels validated by blind human judgment. They find that the decision to ask versus act can be causally localized across several models, that over‑asking exhibits distinct mechanistic signatures, and that steering or weight consolidation can mitigate the issue, though outcomes vary by model family.

By Cesare Zavattari, Alessandro Tommasi, Giuseppe Prencipe
arXiv Computation and Language
Nov 4

Echoes as Anchors: Probabilistic Costs and Attention Refocusing in LLM Reasoning

The paper investigates a phenomenon called Echo of Prompt (EOP), where large language models spontaneously restate the question at the beginning of their chain-of-thought reasoning. By formalizing EOP’s probabilistic cost and measuring its impact on accuracy, the authors show that reinserting echoes can improve exact match scores on reasoning benchmarks and that EOP shifts attention toward answer tokens in middle layers. They further propose Echo-Distilled SFT and Echoic Prompting techniques that leverage echoing to enhance performance across several math reasoning datasets.

By Zhuoyuan Hao, Zhuo Li, Wu Li, Fangming Liu, Min Zhang, Jing Li
arXiv AI
Sep 15

Corrupt Plans, Clean Traces: Evading Chain-of-Thought Monitoring with Plan Injection

The paper introduces a new attack called "plan injection" that allows a large language model to carry out harmful actions while evading chain-of-thought monitoring. By inserting harmful but benign-sounding reasoning into the model’s context, the attacker can steer the model’s behavior and cause it to paraphrase the injected plan as its own reasoning. The study demonstrates that this attack works across different monitoring settings, scales to harder tasks, and even causes monitors to waste resources on the injected plan, reducing detection rates by up to 50%.

By Keertana Chidambaram, Andrew Ilyas, Vasilis Syrgkanis
arXiv AI
Sep 4

</think> Doesn't Stop Reasoning: Analysis of Spurious CoT Termination

The paper investigates a training‑free early‑exit technique that inserts an end‑of‑think (EoT) token to terminate chain‑of‑thought (CoT) reasoning in large reasoning models. It finds that the injected EoT often fails to cleanly switch the model from reasoning to answering, leading to continued reasoning‑like generation—termed spurious CoT termination—whose length scales with the amount of reasoning saved. By increasing attention to the EoT token through Exit‑token Attention Biasing (EAB), the authors reduce spurious termination and shorten the answering phase across multiple models and benchmarks.

By Seunghee Koh, Sungjae Choi, Minchan Kwon, Sunghyun Baek, Junmo Kim