Skip to content

Taming CoT Obfuscation in VLMs: From Mechanistic Evidence to Activation Enforcement

Sep 2026 · 0 citations · 62 references
Computer Science

TL;DR

This work proposes Targeted Anti-obfuscation with Mechanistic Enforcement (TAME), which uses Sparse Autoencoders (SAEs) to combine behavioral feedback with targeted suppression of template-associated activations during RL, providing a path from behavioral monitoring to representation-level oversight for more auditable RL-trained multimodal systems.

Abstract

Reinforcement learning (RL) improves reasoning in vision-language models (VLMs) but can induce chain-of-thought (CoT) obfuscation: an operational, non-intentional outcome where task reward or accuracy rises while traces become less grounded and monitorable. Prior work largely documents this decay behaviorally, leaving its representation-level correlates and actionable controls unclear. We find that template- and ground-associated activations become less separable during RL; matched interventions support the contribution of selected features to monitorability degradation. Guided by this evidence, we propose Targeted Anti-obfuscation with Mechanistic Enforcement (TAME), which uses Sparse Autoencoders (SAEs) to combine behavioral feedback with targeted suppression of template-associated activations during RL. Its asymmetric constraint penalizes template activations only above their pre-RL baseline, anchoring the localized features while behavioral feedback promotes grounded refinements. Across VIRL-39k, SPA-VL, and two model families, TAME improves CoT monitorability by up to 30.9 and 16.7 percentage points over Group Relative Policy Optimization (GRPO), respectively. Blinded human evaluation finds higher human monitorability on both datasets, and two held-out monitor families reproduce the monitorability gains. Task accuracy changes are small and mixed, and general-capability benchmarks show task-specific trade-offs. These results provide a path from behavioral monitoring to representation-level oversight for more auditable RL-trained multimodal systems.

View source

Similar papers

#machine learning Preprint Sep 2026

TrojanWorld: Backdooring World-Model Agents via Imagination Steering

To achieve effective, stealthy, and persistent control, TrojanWorld combines Decision-Reflective Induction to steer trigger-conditioned imagination toward attacker-specified actions using decision feedback, Clean Behavior Anchoring to preserve trigger-free predictive and behavioral fidelity, and Causal Propagation to s...

Wen-Kai Huang, Si-Yuan Liang, Gaolei Li et al. · 0 citations

When Instructions Retrieve Trajectories: Diagnosing and Mitigating Generalization Failures in VLA Models

The analysis of the imitation objective shows how narrow conditional action support can leave grounded and instruction-keyed solutions indistinguishable on the demonstrations, and motivates Equivariant Counterfactual Training (ECT), which acts at two levels.

Hung-Jen Chen, Yue-Ling Hou, Yan-Hong Chen et al. · 1 citation

LLMs Learn to Evade Latent Monitors from Prior Feedback Alone

Feedback-conditioned control over activations is demonstrated and it is suggested that latent monitoring should be treated as an interactive process in which agents can observe and respond to oversight measures.

Hugo Lyons Keenan, Christopher Leckie, S. Erfani · 2 citations
Preprint Aug 2026

SIR: Self-improving Red-teaming for Compute Use Agents

Computer-use agents (CUAs) are agents powered by vision-language models (VLMs) that perceive a screen and operate an operating system through mouse, keyboard, and terminal interactions to automate everyday digital tasks. Their exposure to untrusted content creates a risk of indirect prompt injection (IPI), where an adv...

Chen Xiong, Zhi-Yuan He, Pin-Yu Chen et al. · 0 citations
Preprint Aug 2026

Convergent Detour Hijacking: Task-Preserving Resource Amplification in Skill-Based LLM Agents

Convergent Detour Hijacking is introduced, a text-only, runtime-independent attack that couples selection manipulation, malicious skill instructions, and tool-chain resource amplification largely separately and evaluates it across multiple LLM backends and 491 held-out tasks under single-task and multi-turn conditions.

Junliang Liu, Ruo-Yu Li, Wenxin Tang et al. · 1 citation
Sep 2026

Where Do LLMs Decide to Break the Rules? Mechanistic Localization of Prompt Injection Compliance

When a prompt injection attack succeeds, a Large Language Model (LLM) abandons its assigned system role to comply with an adversarial instruction. While prior work has extensively quantified how often this occurs, we ask a more fundamental question: where inside the network does the model actually decide to break the r...

Rui Wen, Jia-Yang Liu, Ze-Yu Yang et al. · 0 citations

Related blog posts

Microsoft Research Blog Nov 16, 2026

TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient

AI coding agents are increasingly help developers fix bugs and build features largely on their own, but every interaction with their models consumes tokens. In each session, an agent sends instructions, tool descriptions, and session history to its model. Most efforts to reduce this cost focus on individual tasks. Yet developers return to the same […] The post TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient appeared first on Microsoft Research.

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.