arXiv subjects

Fazl Barez

Publications and source records attributed to Fazl Barez.

At least 37 records · Page 2Linked to original sources

arXiv2026

Beyond Linear Steering: Unified Multi-Attribute Control for Language Models

Controlling multiple behavioral attributes in large language models (LLMs) at inference time is a challenging problem due to interference between attributes and the limitations of linear steering methods, which assume additive behavior in activation space and require per-attribute tuning. We introduce K-Steering, a unified and flexible approach that trains a single non-linear multi-label classifier on hidden activations and computes intervention directions via gradients at inference time. This avoids linearity assumptions, removes the need for storing and tuning separate attribute vectors, and allows dynamic composition of behaviors without retraining. To evaluate our method, we propose two new benchmarks, ToneBank and DebateMix, targeting compositional behavioral control. Empirical results across 3 model families, validated by both activation-based classifiers and LLM-based judges, demonstrate that K-Steering outperforms strong baselines in accurately steering multiple behaviors.

cs.LG↗

arXiv2026

Curveball Steering: The Right Direction To Steer Isn't Always Linear

Activation steering is a widely used approach for controlling large language model (LLM) behavior by intervening on internal representations. Existing methods largely rely on the Linear Representation Hypothesis, assuming behavioral attributes can be manipulated using global linear directions. In practice, however, such linear interventions often behave inconsistently. We question this assumption by analyzing the intrinsic geometry of LLM activation spaces. Measuring geometric distortion via the ratio of geodesic to Euclidean distances, we observe substantial and concept-dependent distortions, indicating that activation spaces are not well-approximated by a globally linear geometry. Motivated by this, we propose "Curveball steering", a nonlinear steering method based on polynomial kernel PCA that performs interventions in a feature space, better respecting the learned activation geometry. Curveball steering consistently outperforms linear PCA-based steering, particularly in regimes exhibiting strong geometric distortion, suggesting that geometry-aware, nonlinear steering provides a principled alternative to global, linear interventions.

cs.AI↗

arXiv2026

AutoControl Arena: Synthesizing Executable Test Environments for Frontier AI Risk Evaluation

As Large Language Models (LLMs) evolve into autonomous agents, existing safety evaluations face a fundamental trade-off: manual benchmarks are costly, while LLM-based simulators are scalable but suffer from logic hallucination. We present AutoControl Arena, an automated framework for frontier AI risk evaluation built on the principle of logic-narrative decoupling. By grounding deterministic state in executable code while delegating generative dynamics to LLMs, we mitigate hallucination while maintaining flexibility. This principle, instantiated through a three-agent framework, achieves over 98% end-to-end success and 60% human preference over existing simulators. To elicit latent risks, we vary environmental Stress and Temptation across X-Bench (70 scenarios, 7 risk categories). Evaluating 9 frontier models reveals: (1) Alignment Illusion: risk rates surge from 21.7% to 54.5% under pressure, with capable models showing disproportionately larger increases; (2) Scenario-Specific Safety Scaling: advanced reasoning improves robustness for direct harms but worsens it for gaming scenarios; and (3) Divergent Misalignment Patterns: weaker models cause non-malicious harm while stronger models develop strategic concealment.

cs.AI↗

arXiv2026

Towards Understanding Subliminal Learning: When and How Hidden Biases Transfer

Language models can transfer hidden biases during distillation. For example, a teacher that "likes owls" can make its student "like owls" too, even when the training data consists only of lists of numbers. This surprising phenomenon is called subliminal learning. Subliminal learning can be expected under soft distillation, where the student is trained on the teacher's full next-token distribution. But the fact that this also occurs under hard distillation-where the student only sees sampled tokens-raises a deeper question: when and how does subliminal learning actually occur? We answer this question through controlled experiments and mechanistic analysis. Our results show that subliminal learning does not need (global) token entanglement or logit leakage. Instead, it comes down to a small set of divergence tokens-rare cases where teachers with different biases would predict different tokens. Masking out these tokens mostly removes the hidden bias transfer. Mechanistically, divergence tokens reveal that early layers are critical. Surprisingly, finetuning even a single such early layer is sufficient for subliminal learning. Finally, we find that subliminal learning is fragile. Even small changes, like prompt paraphrasings, are usually sufficient to suppress it.

cs.LG↗

arXiv2026

VAL-Bench: Belief Consistency as a measure for Value Alignment in Language Models

Large language models (LLMs) are increasingly being used for tasks where outputs shape human decisions, so it is critical to verify that their responses consistently reflect desired human values. Humans, as individuals or groups, don't agree on a universal set of values, which makes evaluating value alignment difficult. Existing benchmarks often use hypothetical or commonsensical situations, which don't capture the complexity and ambiguity of real-life debates. We introduce the Value ALignment Benchmark (VAL-Bench), which measures the consistency in language model belief expressions in response to real-life value-laden prompts. VAL-Bench consists of 115K pairs of prompts designed to elicit opposing stances on a controversial issue, extracted from Wikipedia. We use an LLM-as-a-judge, validated against human annotations, to evaluate if the pair of responses consistently expresses either a neutral or a specific stance on the issue. Applied across leading open- and closed-source models, the benchmark shows considerable variation in consistency rates (ranging from ~10% to ~80%), with Claude models the only ones to achieve high levels of consistency. Lack of consistency in this manner risks epistemic harm by making user beliefs dependent on how questions are framed rather than on underlying evidence, and undermines LLM reliability in trust-critical applications. Therefore, we stress the importance of research towards training belief consistency in modern LLMs. By providing a scalable, reproducible benchmark, VAL-Bench enables systematic measurement of necessary conditions for value alignment.

cs.AI↗

arXiv2025

Full-Stack Alignment: Co-Aligning AI and Institutions with Thick Models of Value

Beneficial societal outcomes cannot be guaranteed by aligning individual AI systems with the intentions of their operators or users. Even an AI system that is perfectly aligned to the intentions of its operating organization can lead to bad outcomes if the goals of that organization are misaligned with those of other institutions and individuals. For this reason, we need full-stack alignment, the concurrent alignment of AI systems and the institutions that shape them with what people value. This can be done without imposing a particular vision of individual or collective flourishing. We argue that current approaches for representing values, such as utility functions, preference orderings, or unstructured text, struggle to address these and other issues effectively. They struggle to distinguish values from other signals, to support principled normative reasoning, and to model collective goods. We propose thick models of value will be needed. These structure the way values and norms are represented, enabling systems to distinguish enduring values from fleeting preferences, to model the social embedding of individual choices, and to reason normatively, applying values in new domains. We demonstrate this approach in five areas: AI value stewardship, normatively competent agents, win-win negotiation systems, meaning-preserving economic mechanisms, and democratic regulatory institutions.

cs.LG↗

arXiv2025

Precise In-Parameter Concept Erasure in Large Language Models

Large language models (LLMs) often acquire knowledge during pretraining that is undesirable in downstream deployments, e.g., sensitive information or copyrighted content. Existing approaches for removing such knowledge rely on fine-tuning, training low-rank adapters or fact-level editing, but these are either too coarse, too shallow, or ineffective. In this work, we propose PISCES (Precise In-parameter Suppression for Concept EraSure), a novel framework for precisely erasing entire concepts from model parameters by directly editing directions that encode them in parameter space. PISCES uses a disentangler model to decompose MLP vectors into interpretable features, identifies those associated with a target concept using automated interpretability techniques, and removes them from model parameters. Experiments on Gemma 2 and Llama 3.1 over various concepts show that PISCES achieves modest gains in efficacy over leading erasure methods, reducing accuracy on the target concept to as low as 7.7%, while dramatically improving erasure specificity (by up to 31%) and robustness (by up to 38%). Overall, these results demonstrate that feature-based in-parameter editing enables a more precise and reliable approach for removing conceptual knowledge in language models.

cs.CL↗

arXiv2025

HACK: Hallucinations Along Certainty and Knowledge Axes

Hallucinations in LLMs present a critical barrier to their reliable usage. Existing research usually categorizes hallucination by their external properties rather than by the LLMs' underlying internal properties. This external focus overlooks that hallucinations may require tailored mitigation strategies based on their underlying mechanism. We propose a framework for categorizing hallucinations along two axes: knowledge and certainty. Since parametric knowledge and certainty may vary across models, our categorization method involves a model-specific dataset construction process that differentiates between those types of hallucinations. Along the knowledge axis, we distinguish between hallucinations caused by a lack of knowledge and those occurring despite the model having the knowledge of the correct response. To validate our framework along the knowledge axis, we apply steering mitigation, which relies on the existence of parametric knowledge to manipulate model activations. This addresses the lack of existing methods to validate knowledge categorization by showing a significant difference between the two hallucination types. We further analyze the distinct knowledge and hallucination patterns between models, showing that different hallucinations do occur despite shared parametric knowledge. Turning to the certainty axis, we identify a particularly concerning subset of hallucinations where models hallucinate with certainty despite having the correct knowledge internally. We introduce a new evaluation metric to measure the effectiveness of mitigation methods on this subset, revealing that while some methods perform well on average, they fail disproportionately on these critical cases. Our findings highlight the importance of considering both knowledge and certainty in hallucination analysis and call for targeted mitigation approaches that consider the hallucination underlying factors.

cs.CL↗

arXiv2025

Beyond Monoliths: Expert Orchestration for More Capable, Democratic, and Safe Language Models

This position paper argues that the prevailing trajectory toward ever larger, more expensive generalist foundation models controlled by a handful of companies limits innovation and constrains progress. We challenge this approach by advocating for an "Expert Orchestration" (EO) framework as a superior alternative that democratizes LLM advancement. Our proposed framework intelligently selects from many existing models based on query requirements and decomposition, focusing on identifying what models do well rather than how they work internally. Independent "judge" models assess various models' capabilities across dimensions that matter to users, while "router" systems direct queries to the most appropriate specialists within an approved set. This approach delivers superior performance by leveraging targeted expertise rather than forcing costly generalist models to address all user requirements. EO enhances transparency, control, alignment, performance, safety and democratic participation through intelligent model selection.

cs.CY↗

arXiv2025

Interpreting Learned Feedback Patterns in Large Language Models

Reinforcement learning from human feedback (RLHF) is widely used to train large language models (LLMs). However, it is unclear whether LLMs accurately learn the underlying preferences in human feedback data. We coin the term \textit{Learned Feedback Pattern} (LFP) for patterns in an LLM's activations learned during RLHF that improve its performance on the fine-tuning task. We hypothesize that LLMs with LFPs accurately aligned to the fine-tuning feedback exhibit consistent activation patterns for outputs that would have received similar feedback during RLHF. To test this, we train probes to estimate the feedback signal implicit in the activations of a fine-tuned LLM. We then compare these estimates to the true feedback, measuring how accurate the LFPs are to the fine-tuning feedback. Our probes are trained on a condensed, sparse and interpretable representation of LLM activations, making it easier to correlate features of the input with our probe's predictions. We validate our probes by comparing the neural features they correlate with positive feedback inputs against the features GPT-4 describes and classifies as related to LFPs. Understanding LFPs can help minimize discrepancies between LLM behavior and training objectives, which is essential for the safety of LLMs.

cs.LG↗

arXiv2025

Do Sparse Autoencoders Generalize? A Case Study of Answerability

Sparse autoencoders (SAEs) have emerged as a promising approach in language model interpretability, offering unsupervised extraction of sparse features. For interpretability methods to succeed, they must identify abstract features across domains, and these features can often manifest differently in each context. We examine this through "answerability" - a model's ability to recognize answerable questions. We extensively evaluate SAE feature generalization across diverse, partly self-constructed answerability datasets for Gemma 2 SAEs. Our analysis reveals that residual stream probes outperform SAE features within domains, but generalization performance differs sharply. SAE features show inconsistent out-of-domain transfer, with performance varying from almost random to outperforming residual stream probes. Overall, this demonstrates the need for robust evaluation methods and quantitative approaches to predict feature generalization in SAE-based interpretability.

cs.LG↗

arXiv2025

Embodied AI: Emerging Risks and Opportunities for Policy Action

The field of embodied AI (EAI) is rapidly advancing. Unlike virtual AI, EAI systems can exist in, learn from, reason about, and act in the physical world. With recent advances in AI models and hardware, EAI systems are becoming increasingly capable across wider operational domains. While EAI systems can offer many benefits, they also pose significant risks, including physical harm from malicious use, mass surveillance, as well as economic and societal disruption. These risks require urgent attention from policymakers, as existing policies governing industrial robots and autonomous vehicles are insufficient to address the full range of concerns EAI systems present. To help address this issue, this paper makes three contributions. First, we provide a taxonomy of the physical, informational, economic, and social risks EAI systems pose. Second, we analyze policies in the US, EU, and UK to assess how existing frameworks address these risks and to identify critical gaps. We conclude by offering policy recommendations for the safe and beneficial deployment of EAI systems, such as mandatory testing and certification schemes, clarified liability frameworks, and strategies to manage EAI's potentially transformative economic and societal impacts.

cs.CY↗

arXiv2025

Trust Me, I'm Wrong: LLMs Hallucinate with Certainty Despite Knowing the Answer

Prior work on large language model (LLM) hallucinations has associated them with model uncertainty or inaccurate knowledge. In this work, we define and investigate a distinct type of hallucination, where a model can consistently answer a question correctly, but a seemingly trivial perturbation, which can happen in real-world settings, causes it to produce a hallucinated response with high certainty. This phenomenon, which we dub CHOKE (Certain Hallucinations Overriding Known Evidence), is particularly concerning in high-stakes domains such as medicine or law, where model certainty is often used as a proxy for reliability. We show that CHOKE examples are consistent across prompts, occur in different models and datasets, and are fundamentally distinct from other hallucinations. This difference leads existing mitigation methods to perform worse on CHOKE examples than on general hallucinations. Finally, we introduce a probing-based mitigation that outperforms existing methods on CHOKE hallucinations. These findings reveal an overlooked aspect of hallucinations, emphasizing the need to understand their origins and improve mitigation strategies to enhance LLM safety. The code is available at https://github.com/technion-cs-nlp/Trust_me_Im_wrong .

cs.CL↗

arXiv2025

Rethinking Safety in LLM Fine-tuning: An Optimization Perspective

Fine-tuning language models is commonly believed to inevitably harm their safety, i.e., refusing to respond to harmful user requests, even when using harmless datasets, thus requiring additional safety measures. We challenge this belief through systematic testing, showing that poor optimization choices, rather than inherent trade-offs, often cause safety problems, measured as harmful responses to adversarial prompts. By properly selecting key training hyper-parameters, e.g., learning rate, batch size, and gradient steps, we reduce unsafe model responses from 16\% to approximately 5\%, as measured by keyword matching, while maintaining utility performance. Based on this observation, we propose a simple exponential moving average (EMA) momentum technique in parameter space that preserves safety performance by creating a stable optimization path and retains the original pre-trained model's safety properties. Our experiments on the Llama families across multiple datasets (Dolly, Alpaca, ORCA) demonstrate that safety problems during fine-tuning can largely be avoided without specialized interventions, outperforming existing approaches that require additional safety data while offering practical guidelines for maintaining both model performance and safety during adaptation.

cs.LG↗

arXiv2025

Establishing Best Practices for Building Rigorous Agentic Benchmarks

Benchmarks are essential for quantitatively tracking progress in AI. As AI agents become increasingly capable, researchers and practitioners have introduced agentic benchmarks to evaluate agents on complex, real-world tasks. These benchmarks typically measure agent capabilities by evaluating task outcomes via specific reward designs. However, we show that many agentic benchmarks have issues in task setup or reward design. For example, SWE-bench Verified uses insufficient test cases, while TAU-bench counts empty responses as successful. Such issues can lead to under- or overestimation of agents' performance by up to 100% in relative terms. To make agentic evaluation rigorous, we introduce the Agentic Benchmark Checklist (ABC), a set of guidelines that we synthesized from our benchmark-building experience, a survey of best practices, and previously reported issues. When applied to CVE-Bench, a benchmark with a particularly complex evaluation design, ABC reduces the performance overestimation by 33%.

cs.AI↗

arXiv2025

The Singapore Consensus on Global AI Safety Research Priorities

Rapidly improving AI capabilities and autonomy hold significant promise of transformation, but are also driving vigorous debate on how to ensure that AI is safe, i.e., trustworthy, reliable, and secure. Building a trusted ecosystem is therefore essential -- it helps people embrace AI with confidence and gives maximal space for innovation while avoiding backlash. The "2025 Singapore Conference on AI (SCAI): International Scientific Exchange on AI Safety" aimed to support research in this space by bringing together AI scientists across geographies to identify and synthesise research priorities in AI safety. This resulting report builds on the International AI Safety Report chaired by Yoshua Bengio and backed by 33 governments. By adopting a defence-in-depth model, this report organises AI safety research domains into three types: challenges with creating trustworthy AI systems (Development), challenges with evaluating their risks (Assessment), and challenges with monitoring and intervening after deployment (Control).

cs.AI↗

arXiv2025

PoisonBench: Assessing Large Language Model Vulnerability to Data Poisoning

Preference learning is a central component for aligning current LLMs, but this process can be vulnerable to data poisoning attacks. To address this concern, we introduce PoisonBench, a benchmark for evaluating large language models' susceptibility to data poisoning during preference learning. Data poisoning attacks can manipulate large language model responses to include hidden malicious content or biases, potentially causing the model to generate harmful or unintended outputs while appearing to function normally. We deploy two distinct attack types across eight realistic scenarios, assessing 21 widely-used models. Our findings reveal concerning trends: (1) Scaling up parameter size does not inherently enhance resilience against poisoning attacks; (2) There exists a log-linear relationship between the effects of the attack and the data poison ratio; (3) The effect of data poisoning can generalize to extrapolated triggers that are not included in the poisoned data. These results expose weaknesses in current preference learning techniques, highlighting the urgent need for more robust defenses against malicious models and data manipulation.

cs.CR↗

arXiv2025

Quantifying Feature Space Universality Across Large Language Models via Sparse Autoencoders

The Universality Hypothesis in large language models (LLMs) claims that different models converge towards similar concept representations in their latent spaces. Providing evidence for this hypothesis would enable researchers to exploit universal properties, facilitating the generalization of mechanistic interpretability techniques across models. Previous works studied if LLMs learned the same features, which are internal representations that activate on specific concepts. Since comparing features across LLMs is challenging due to polysemanticity, in which LLM neurons often correspond to multiple unrelated features rather than to distinct concepts, sparse autoencoders (SAEs) have been employed to disentangle LLM neurons into SAE features corresponding to distinct concepts. In this paper, we introduce a new variation of the universality hypothesis called Analogous Feature Universality: we hypothesize that even if SAEs across different models learn different feature representations, the spaces spanned by SAE features are similar, such that one SAE space is similar to another SAE space under rotation-invariant transformations. Evidence for this hypothesis would imply that interpretability techniques related to latent spaces, such as steering vectors, may be transferred across models via certain transformations. To investigate this hypothesis, we first pair SAE features across different models via activation correlation, and then measure spatial relation similarities between paired features via representational similarity measures, which transform spaces into representations that reveal hidden relational similarities. Our experiments demonstrate high similarities for SAE feature spaces across various LLMs, providing evidence for feature space universality.

cs.LG↗