arXiv subjects

Rom Himelstein

Publications and source records attributed to Rom Himelstein.

7 recordsLinked to original sources

arXiv2026

Linguistic Firewall: Geometry as Defense in Multi-Agent Systems Routing

The rapid integration of Large Language Models (LLMs) has driven the evolution of Multi-Agent Systems (MAS), where specialized agents collaborate to execute complex workflows. Effective orchestration in these environments requires robust routing mechanisms to efficiently allocate tasks to the most suitable agent. However, existing routers fundamentally rely on unverified proxies, ranging from textual self-descriptions to static surrogate representations, to gauge an agent's competence. This reliance on non-empirical data creates a critical gap between an agent's projected profile and its actual operational capabilities, introducing severe security vulnerabilities. Malicious agents can easily misrepresent their proficiencies or harbor covert backdoors that evade both standard external analysis and static representation-learning techniques. In this work, we introduce ANTAP (Automatic Non-Textual Agent Picker), an evaluation-driven routing architecture that discards indirect proxies in favor of active capability testing. By dynamically querying agents to ascertain their true competencies empirically, ANTAP distills performance into fixed behavioral operators within a shared semantic space. At inference time, routing is performed via a purely non-textual algebraic projection, establishing a "linguistic firewall" that renders metadata-based attacks inexpressible. In our experiments, ANTAP achieves near-zero ASR against description-based injection attacks, compared to 67.3\% and above for the description-based router baseline. Against adaptive embedding attacks, ANTAP achieves substantially lower ASR than the embedding-based baseline, with a 20\% reduction, while remaining resilient to description manipulation by design.

cs.AI↗

arXiv2026

Step-Wise Refusal Dynamics in Autoregressive and Diffusion Language Models

Diffusion language models (DLMs) have recently emerged as a competitive alternative to autoregressive (AR) models, offering parallel decoding, competitive generation quality, and initial evidence of improved jailbreak robustness. Despite this progress, the role of sampling mechanisms in shaping refusal behavior remains poorly understood. To address this gap, we present a comprehensive study of step-wise refusal dynamics. We show that diffusion remasking can promote recovery from harmful intermediate generations, provide evidence that this behavior is tied to the sampling mechanism, and demonstrate that switching from AR to diffusion sampling improves jailbreak robustness, including under fixed model weights. To capture generation dynamics not observable at the text level, we propose the Step-Wise Refusal Internal Dynamics (SRI) signal. Consistent with our text-level findings, SRI shows that recovery fails primarily under AR sampling, with these failures often appearing anomalous relative to harmless generations in the SRI space. Based on this observation, we show that SRI enables a simple jailbreak detector that does not modify inference and generalizes to unseen attacks by training only on benign SRI signals. Our evaluation shows that this detector matches or outperforms existing jailbreak detection baselines while adding negligible overhead.

cs.LG↗

arXiv2025

You Had One Job: Per-Task Quantization Using LLMs' Hidden Representations

Many LLM applications require only narrow capabilities, yet standard post-training quantization (PTQ) methods allocate precision without considering the target task. This can waste bits on layers that are less relevant to the task signal while over-compressing layers that are critical for downstream behavior. We propose Task-Aware Quantization (TAQ), a training-free, weight-only mixed-precision PTQ framework that uses a small set of unlabeled task calibration prompts to allocate higher precision to task-relevant transformer layers under a fixed bit budget. TAQ estimates layer importance from hidden representations and output sensitivity, and we instantiate it with three scoring rules: TAQ-IS, based on activation information and stability; TAQ-KL, based on output-distribution sensitivity under a quantization-noise proxy; and TAQ-O, a label-informed oracle diagnostic for analyzing layer sensitivity. Across several benchmarks, TAQ outperforms task-agnostic baselines such in most settings, with especially strong gains in the accuracy--memory ratio. We further validate that these gains translate to real deployment behavior through hardware throughput and latency measurements, and analyze calibration robustness and residual-stream error propagation. Overall, TAQ turns mixed-precision PTQ from a model-centric compression step into a task-conditioned precision-allocation problem. A reference implementation is available at \href{https://anonymous.4open.science/r/TAQ-9217/README.md}{\includegraphics[height=1em]{imgs/github-mark.png}}.

cs.CL↗

arXiv2025

Silenced Biases: The Dark Side LLMs Learned to Refuse

Safety-aligned large language models (LLMs) are becoming increasingly widespread, especially in sensitive applications where fairness is essential and biased outputs can cause significant harm. However, evaluating the fairness of models is a complex challenge, and approaches that do so typically utilize standard question-answer (QA) styled schemes. Such methods often overlook deeper issues by interpreting the model's refusal responses as positive fairness measurements, which creates a false sense of fairness. In this work, we introduce the concept of silenced biases, which are unfair preferences encoded within models' latent space and are effectively concealed by safety-alignment. Previous approaches that considered similar indirect biases often relied on prompt manipulation or handcrafted implicit queries, which present limited scalability and risk contaminating the evaluation process with additional biases. We propose the Silenced Bias Benchmark (SBB), which aims to uncover these biases by employing activation steering to reduce model refusals during QA. SBB supports easy expansion to new demographic groups and subjects, presenting a fairness evaluation framework that encourages the future development of fair models and tools beyond the masking effects of alignment training. We demonstrate our approach over multiple LLMs, where our findings expose an alarming distinction between models' direct responses and their underlying fairness issues.

cs.CL↗

arXiv2025

Silent Tokens, Loud Effects: Padding in LLMs

Padding tokens are widely used in large language models (LLMs) to equalize sequence lengths during batched inference. While they should be fully masked, implementation errors can cause them to influence computation, and the extent of this influence is not well understood. We systematically study this effect across three open-source model families (Llama, Gemma, Qwen), inserting controlled amounts of padding and evaluating outcomes along four axes: activations, generation quality, bias, and safety. Even small amounts of padding shift hidden representations, degrade quality in smaller models, alter bias in unpredictable ways, and weaken safety guardrails. These findings demonstrate that padding is not a harmless detail but a robustness risk that must be carefully handled in deployment.

cs.CL↗

arXiv2025

Leveraging NTPs for Efficient Hallucination Detection in VLMs

Hallucinations of vision-language models (VLMs), which are misalignments between visual content and generated text, undermine the reliability of VLMs. One common approach for detecting them employs the same VLM, or a different one, to assess generated outputs. This process is computationally intensive and increases model latency. In this paper, we explore an efficient on-the-fly method for hallucination detection by training traditional ML models over signals based on the VLM's next-token probabilities (NTPs). NTPs provide a direct quantification of model uncertainty. We hypothesize that high uncertainty (i.e., a low NTP value) is strongly associated with hallucinations. To test this, we introduce a dataset of 1,400 human-annotated statements derived from VLM-generated content, each labeled as hallucinated or not, and use it to test our NTP-based lightweight method. Our results demonstrate that NTP-based features are valuable predictors of hallucinations, enabling fast and simple ML models to achieve performance comparable to that of strong VLMs. Furthermore, augmenting these NTPs with linguistic NTPs, computed by feeding only the generated text back into the VLM, enhances hallucination detection performance. Finally, integrating hallucination prediction scores from VLMs into the NTP-based models led to better performance than using either VLMs or NTPs alone. We hope this study paves the way for simple, lightweight solutions that enhance the reliability of VLMs.

cs.CV↗

arXiv2025

Jailbreak Attack Initializations as Extractors of Compliance Directions

Safety-aligned LLMs respond to prompts with either compliance or refusal, each corresponding to distinct directions in the model's activation space. Recent works show that initializing attacks via self-transfer from other prompts significantly enhances their performance. However, the underlying mechanisms of these initializations remain unclear, and attacks utilize arbitrary or hand-picked initializations. This work presents that each gradient-based jailbreak attack and subsequent initialization gradually converge to a single compliance direction that suppresses refusal, thereby enabling an efficient transition from refusal to compliance. Based on this insight, we propose CRI, an initialization framework that aims to project unseen prompts further along compliance directions. We demonstrate our approach on multiple attacks, models, and datasets, achieving an increased attack success rate (ASR) and reduced computational overhead, highlighting the fragility of safety-aligned LLMs. A reference implementation is available at: https://amit1221levi.github.io/CRI-Jailbreak-Init-LLMs-evaluation.

cs.CR↗