arXiv subjects

Uzay Macar

Publications and source records attributed to Uzay Macar.

8 recordsLinked to original sources

arXiv2026

Measuring Activation Control in Large Language Models

Safe deployment of increasingly capable models will likely come to rely on latent-space monitoring as a complement to behavioral evaluations, especially when evaluation-aware models exhibit scheming or deception. However, if models can also control their own activations, deception could extend into the latent space itself. With this in mind, we introduce the Activation Controllability Benchmark to quantify the extent to which models can modulate their residual stream via natural-language instruction. Across model families and capability levels, we find that most LLMs can control the direction and magnitude of their residual stream activations with some degree of temporal resolution, though performance varies considerably across models. In simple tasks, this level of control can evade activation-based monitoring methods (including linear probes, natural language autoencoders, activation oracles, and the Jacobian lens), albeit imperfectly. These results suggest that control over the activation space itself could become a confound for monitoring as introspective capabilities increase; therefore, we recommend that frontier labs and evaluators track activation controllability in future models.

cs.AI↗

arXiv2026

Mechanisms of Introspective Awareness

Recent work has shown that LLMs can sometimes detect when steering vectors are injected into their residual stream and identify the injected concept -- a phenomenon termed "introspective awareness." We investigate the mechanisms underlying this capability in open-weights models. First, we find that it is behaviorally robust: models detect injected steering vectors at moderate rates with 0% false positives across diverse prompts and dialogue formats. Notably, this capability emerges specifically from post-training; we show that preference optimization algorithms like DPO can elicit it, but standard supervised finetuning does not. We provide evidence that detection cannot be explained by simple linear association between certain steering vectors and directions promoting affirmative responses. We trace the detection mechanism to a two-stage circuit in which "evidence carrier" features in early post-injection layers detect perturbations monotonically along diverse directions, suppressing downstream "gate" features that implement a default negative response. This circuit is absent in base models and robust to refusal ablation. Identification of injected concepts relies on largely distinct later-layer mechanisms that only weakly overlap with those involved in detection. Finally, we show that introspective capability is substantially underelicited: ablating refusal directions improves detection by +53%, and a trained bias vector improves it by +75% on held-out concepts, both without meaningfully increasing false positives. Our results suggest that this introspective awareness of injected concepts is robust and mechanistically nontrivial, and could be substantially amplified in future models. Code: https://github.com/safety-research/introspection-mechanisms.

cs.LG↗

arXiv2025

Thought Branches: Interpreting LLM Reasoning Requires Resampling

Most work interpreting reasoning models studies only a single chain-of-thought (CoT), yet these models define distributions over many possible CoTs. We argue that studying a single sample is inadequate for understanding causal influence and the underlying computation. Though fully specifying this distribution is intractable, we can measure a partial CoT's impact by resampling only the subsequent text. We present case studies using resampling to investigate model decisions. First, when a model states a reason for its action, does that reason actually cause the action? In "agentic misalignment" scenarios, we find that self-preservation sentences have small causal impact, suggesting they do not meaningfully drive blackmail. Second, are artificial edits to CoT sufficient for steering reasoning? Resampling and selecting a completion with the desired property is a principled on-policy alternative. We find that off-policy interventions yield small and unstable effects compared to resampling in decision-making tasks. Third, how do we understand the effect of removing a reasoning step when the model may repeat it post-edit? We introduce a resilience metric that repeatedly resamples to prevent similar content from reappearing downstream. Critical planning statements resist removal but have large effects when eliminated. Fourth, since CoT is sometimes "unfaithful", can our methods teach us anything in these settings? Adapting causal mediation analysis, we find that hints that causally affect the output without being explicitly mentioned exert a subtle and cumulative influence on the CoT that persists even if the hint is removed. Overall, studying distributions via resampling enables reliable causal analysis, clearer narratives of model reasoning, and principled CoT interventions.

cs.LG↗

arXiv2025

Thought Anchors: Which LLM Reasoning Steps Matter?

Current frontier large-language models rely on reasoning to achieve state-of-the-art performance. Many existing interpretability are limited in this area, as standard methods have been designed to study single forward passes of a model rather than the multi-token computational steps that unfold during reasoning. We argue that analyzing reasoning traces at the sentence level is a promising approach to understanding reasoning processes. We introduce a black-box method that measures each sentence's counterfactual importance by repeatedly sampling replacement sentences from the model, filtering for semantically different ones, and continuing the chain of thought from that point onwards to quantify the sentence's impact on the distribution of final answers. We discover that certain sentences can have an outsized impact on the trajectory of the reasoning trace and final answer. We term these sentences \textit{thought anchors}. These are generally planning or uncertainty management sentences, and specialized attention heads consistently attend from subsequent sentences to thought anchors. We further show that examining sentence-sentence causal links within a reasoning trace gives insight into a model's behavior. Such information can be used to predict a problem's difficulty and the extent different question domains involve sequential or diffuse reasoning. As a proof-of-concept, we demonstrate that our techniques together provide a practical toolkit for analyzing reasoning models by conducting a detailed case study of how the model solves a difficult math problem, finding that our techniques yield a consistent picture of the reasoning trace's structure. We provide an open-source tool (thought-anchors.com) for visualizing the outputs of our methods on further problems. The convergence across our methods shows the potential of sentence-level analysis for a deeper understanding of reasoning models.

cs.LG↗

arXiv2025

LogicLearner: A Tool for the Guided Practice of Propositional Logic Proofs

The study of propositional logic -- fundamental to the theory of computing -- is a cornerstone of the undergraduate computer science curriculum. Learning to solve logical proofs requires repeated guided practice, but undergraduate students often lack access to on-demand tutoring in a judgment-free environment. In this work, we highlight the need for guided practice tools in undergraduate mathematics education and outline the desiderata of an effective practice tool. We accordingly develop LogicLearner, a web application for guided logic proof practice. LogicLearner consists of an interface to attempt logic proofs step-by-step and an automated proof solver to generate solutions on the fly, allowing users to request guidance as needed. We pilot LogicLearner as a practice tool in two semesters of an undergraduate discrete mathematics course and receive strongly positive feedback for usability and pedagogical value in student surveys. To the best of our knowledge, LogicLearner is the only learning tool that provides an end-to-end practice environment for logic proofs with immediate, judgment-free feedback.

cs.DM↗

arXiv2024

Hierarchical Multi-Armed Bandits for the Concurrent Intelligent Tutoring of Concepts and Problems of Varying Difficulty Levels

Remote education has proliferated in the twenty-first century, yielding rise to intelligent tutoring systems. In particular, research has found multi-armed bandit (MAB) intelligent tutors to have notable abilities in traversing the exploration-exploitation trade-off landscape for student problem recommendations. Prior literature, however, contains a significant lack of open-sourced MAB intelligent tutors, which impedes potential applications of these educational MAB recommendation systems. In this paper, we combine recent literature on MAB intelligent tutoring techniques into an open-sourced and simply deployable hierarchical MAB algorithm, capable of progressing students concurrently through concepts and problems, determining ideal recommended problem difficulties, and assessing latent memory decay. We evaluate our algorithm using simulated groups of 500 students, utilizing Bayesian Knowledge Tracing to estimate students' content mastery. Results suggest that our algorithm, when turned difficulty-agnostic, significantly boosts student success, and that the further addition of problem-difficulty adaptation notably improves this metric.

cs.CY↗

arXiv2023

Teenagers and Artificial Intelligence: Bootcamp Experience and Lessons Learned

Artificial intelligence (AI) stands out as a game-changer in today's technology landscape. However, the integration of AI education in classroom curricula currently lags behind, leaving teenagers inadequately prepared for an imminent AI-driven future. In this pilot study, we designed a three-day bootcamp offered in the summer of 2023 to a cohort of 60 high school students. The curriculum was delivered in person through animated video content, easy-to-follow slides, interactive playgrounds, and quizzes. These were packaged in the early version of an online learning platform we are developing. Results from the post-bootcamp survey conveyed a 91.4% overall satisfaction. Despite the short bootcamp duration, 88.5% and 71.4% of teenagers responded that they had an improved understanding of AI concepts and programming, respectively. Overall, we found that employing diverse modalities effectively engaged students, and building foundational modules proved beneficial for introducing more complex topics. Furthermore, using Google Colab notebooks for coding assignments proved challenging to most students. Students' activity on the platform and their answers to quizzes showed proficient engagement and a grasp of the material. Our results strongly highlight the need for compelling and accessible AI education methods for the next generation and the potential for informal learning to fill the gap of providing early AI education to teenagers.

cs.CY↗

arXiv2021

Team NeuroPoly: Description of the Pipelines for the MICCAI 2021 MS New Lesions Segmentation Challenge

This paper gives a detailed description of the pipelines used for the 2nd edition of the MICCAI 2021 Challenge on Multiple Sclerosis Lesion Segmentation. An overview of the data preprocessing steps applied is provided along with a brief description of the pipelines used, in terms of the architecture and the hyperparameters. Our code for this work can be found at: https://github.com/ivadomed/ms-challenge-2021.

eess.IV↗