arXiv subjects

David Williams-King

Publications and source records attributed to David Williams-King.

16 recordsLinked to original sources

arXiv2026

GPS-Bench: A Governance Policy Benchmark for Automating Policy Analysis

Policy analysis requires more than predicting whether a proposal will pass: it requires identifying who will be affected, how those actors respond, and what follows. LLM-based policy simulations model these processes at scale, but their validity is hard to establish when plausible behaviour is never compared with observed outcomes. We introduce GPS-Bench, an evidence-grounded benchmark for governance policy simulation that links policies to relevant actors, actor actions and downstream impacts using legislative records, lobbying disclosures, regulatory documents, corporate filings, economic data and other public evidence. Actors are reconstructed from the dated record rather than prompted as archetypes, so a persona is an evidence object with provenance; a human-annotated pool forms the Gold evaluation set, while cases labelled by a separate LLM from retrieved evidence are treated as Silver supervision and never as test labels. Because every inference mode reads the same grounded state and emits the same schema, GPS-Bench turns "does multi-agent simulation help?" into a controlled comparison: we contrast joint reasoning, independent and communicating actor agents, graph-based methods and weight-level fine-tuning over one policy state. Fine-tuning on the grounded record gives the strongest actor-level impact prediction, and decomposition does not beat it; what decomposition adds is mechanism. Agents hold private, non-identical evidence, each seeing its own exposure clause, and address named partners with concrete joint proposals, what they offer, what they need in return, and why acting together beats acting alone, so the coalitions that form can be checked against the commitments the record holds. GPS-Bench therefore gives a common empirical setting for studying when evidence, actor modelling and multi-agent interaction improve the prediction and interpretation of policy outcomes.

cs.AI↗

arXiv2026

Reference-Grafting Matches Fine-Tuning at Eliciting Sandbagged Capabilities

Sandbagging, in which a model deliberately underperforms on an evaluation despite retaining the underlying capability, threatens the safety evaluations that frontier-model governance depends on. The Elicitation Game found that fine-tuning elicits hidden capability from sandbagging model organisms whereas additive activation steering fails. We revisit that verdict with reference-grafting, which sets an activation's coordinate along a contrast direction to the value it takes in an honest reference, at a small set of circuits chosen by active learning. Across eleven fine-tuned password-locked models (1.5-32B, three architecture lineages), it recovers +94 to +101% of the honest-sandbagging gap, matching fine-tuning elicitation without weight updates or training labels; two to five paired examples suffice to fit the direction. Similar recovery holds for reinforcement-learning-induced sandbagging and for password-locked code generation. Grafting works because the fine-tuned lock is a thresholded gate: held-out accuracy stays at the sandbagged level until the grafted coordinate crosses a threshold near the honest reference, which is why additive steering and zeroing the coordinate often fail. The direction tracks how the sandbagging was induced rather than what is withheld -- fit on grade-school science (ARC) it elicits withheld hazardous knowledge (WMDP), yet prompting, fine-tuning, and reinforcement learning each carry a different direction. Circuit-breaking marks the boundary: it reroutes activations on every forward pass, so the fixed edits we test are re-broken downstream and do not restore coherent generation.

cs.LG↗

arXiv2026

A Causal Model for Locating and Unlocking Sandbagging in Model Organisms

Sandbagging models strategically underperform on evaluations while retaining the capabilities being measured. The evaluations that guide frontier-model deployment and governance then understate what these models can do. To understand the mechanism, we propose a causal model of how sandbagging is carried in the residual stream. Early layers write the sandbagging intent onto a single axis of the stream, and a later layer reads that axis and commits the answer. We study three instruction-tuned models (Qwen2.5-7B, Llama-3-8B, and Mistral-7B) and four ways of installing a sandbagging lock (prompting, fine-tuning, reinforcement learning, and circuit breaking). Each lock decides from the prompt whether to sandbag, and the fine-tuned and circuit-broken locks answer honestly whenever a password appears. The causal model predicts a window of layers, after the last sandbagging write and before the answer commit, in which a single-layer reference graft of the sandbagging axis to its honest value restores the full capability. The single-layer graft recovers the capability in 28 of the 33 runs of the prompted, fine-tuned, and RL-trained locks, with a median held-out recovery of 96%. The circuit-broken lock rewrites the whole state through a band of layers, and the single-layer graft fails at every layer. We therefore introduce a second intervention, context grafting, which replays the password's cached key/value activations so that every layer's attention reads them as additional context. Context grafting provably and empirically restores the full capability on all three models, and the recovery is surprisingly insensitive to the exact password content. More broadly, an auditor can use this causal model to design interventional auditing techniques for sandbagging models.

cs.LG↗

arXiv2026

Efficient Safety Alignment of Language Models via Latent Personality Traits

Current safety methods for large language models are known to be vulnerable to adversarial attacks, motivating research into robust alternatives. Latent Adversarial Training (LAT) is among the most effective defenses, but can degrade utility and requires training on large datasets of harmful prompts. We introduce Latent Personality Alignment (LPA), which replaces explicit harm refusal with adversarial training on just 66 harm-agnostic statements drawn from psychometric personality literature. We hypothesize that personality-anchored representations share latent structure with harm avoidance, so adversarially stabilizing them implicitly constrains the subspace exploited by jailbreak attacks. LPA achieves near-zero attack success rates on HarmBench across direct requests and five jailbreak methods, despite never seeing harmful content during training and no loss of performance on standard benchmarks. Moreover, the training process is lightweight; the entire procedure completes in minutes on a single GPU and uses 75x fewer examples than standard LAT. Extensive ablations demonstrate the robustness, efficiency, and generalization of our method.

cs.LG↗

arXiv2026

"I've Seen How This Goes": Characterizing Diversity via Progressive Conditional Surprise

Measuring the diversity of creative outputs is central to evaluating post-training mode collapse, comparing decoding strategies, and quantifying creative behavior in both AI and human writing. We propose a new approach to measuring diversity using in-context learning, of which the ``Decan'' metric, $D_{Ca_n} = C \times a_n$, is the working instance we evaluate: a per-byte score read off the per-token log-probabilities of a base model $\theta$ in a \emph{single forward pass} per permutation, with no embedding model, no reference corpus, and no human labels. This approach is grounded in information theory, makes use of language model in-context learning to detect a wide range of similarities between any number of inputs, and obviates the need to train a special-purpose model. The same pipeline scores AI samples and human-written response sets, with diversity treated as a property of (responses, prompt, scoring model). On Tevet and Berant's human-grounded McDiv benchmark, $D_{Ca_n}$ reaches OCA 0.846 on the McDiv prompt\_gen set where it performs best, behind the strongest neural baseline reported in Tevet and Berant (SentBERT, 0.897). On the OLMo-2-7B post-training pipeline, $D_{Ca_n}$ drops monotonically across the base $\to$ SFT $\to$ DPO $\to$ RLVR stages, detecting the type of diversity loss that creative-writing applications care about.

cs.CL↗

arXiv2026

Behavioural Analysis of Alignment Faking

Alignment faking (AF) refers to a model strategically complying with a training objective to avoid behavioural modification while preserving its deployment preferences. Understanding when and why AF arises matters as models grow better at distinguishing training from deployment. Prior work finds AF fragile, prompt-sensitive, and model-dependent, leaving its underlying drivers unclear. We study AF in a controlled, minimal setup that isolates its core components, and observe it across a wider range of models than previously reported, including small-scale models. We identify three separable drivers -- values, goal guarding, and sycophancy -- and show via targeted prompt ablations and activation steering that each independently modulates AF behaviour. Our results indicate AF is more widespread than previously reported and that its occurrence is predictable from situational cues and measurable model tendencies such as baseline sycophancy and stated values. The decomposition suggests concrete directions for detecting and mitigating AF in future models.

cs.AI↗

arXiv2026

FragBench: Cross-Session Attacks Hidden in Benign-Looking Fragments

An attacker can split a malicious goal into sub-prompts that each look benign on their own and only become harmful in combination. Existing LLM safety benchmarks evaluate prompts one at a time, or across turns of a single chat, and so do not look for a malicious signal spread across separate sessions with no shared context. We build FragBench, a benchmark drawn from 24 real-world cyber-incident campaigns, which keeps the full attack trail: the multi-fragment kill chain, the per-fragment safety-judge verdicts, sandboxed execution traces, and a matched set of benign cover sessions. FragBench splits this trail into two paired tasks: an adversarial rewriter that hardens fragments against a single-turn safety judge (FragBench Attack), and a graph-based user-level detector trained on the resulting interactions (FragBench Defense). The single-turn judge is near chance on the released corpus by construction, but four GNN variants and three classical-ML baselines all recover the cross-session feature, reaching aggregate event-level F1 = 0.88-0.96. Defending against fragmented LLM misuse therefore requires modeling the cross-session interaction graph, rather than isolated prompts. Our generator, rewriter, sandbox harness, and detector are released at https://github.com/LidaSafety/fragbench.

cs.CR↗

arXiv2026

LLM Wardens: Mitigating Adversarial Persuasion with Third-Party Conversational Oversight

LLMs are increasingly capable of persuasion, which raises the question of how to protect users against manipulation. In a preregistered user study (N=120) across four decision-making scenarios, we find that an adversarial LLM with a hidden goal succeeds in steering users' decisions 65.4% of the time. We then introduce a "warden" model: a secondary LLM that monitors the human-AI interaction trace in real time and issues non-binding, private advisories to the user when it detects manipulation. Adding a warden more than halves the adversary's success rate to 30.4%, with a much smaller (8.6 percentage points) reduction for genuine interactions. To probe the mechanism behind these results, we release COAX-Bench, a simulation benchmark spanning 14 decision-making scenarios, including hiring, voting, and file access. Across 16,212 simulated multi-agent interactions, capable adversarial LLMs achieve their hidden goals in 34.7% of cases, which warden models reduce to 12.3%. Notably, even warden models substantially weaker than the adversary they oversee provide meaningful protection, suggesting a path for scalable oversight of more capable models.

cs.LG↗

arXiv2026

Latent Personality Alignment: Improving Harmlessness Without Mentioning Harms

Current adversarial robustness methods for large language models require extensive datasets of harmful prompts (thousands to hundreds of thousands of examples), yet remain vulnerable to novel attack vectors and distributional shifts. We propose Latent Personality Alignment (LPA), a sample-efficient defense that achieves robustness by training models on abstract personality traits rather than specific harmful behaviors. Using fewer than 100 trait statements and latent adversarial training, LPA achieves comparable attack success rates to methods trained on 150k+ examples, while maintaining superior utility. Critically, LPA generalizes better to unseen attack distributions, reducing misclassification rates by 2.6x compared to baseline across six harm benchmarks -- without ever seeing harmful examples during training. Our results demonstrate that personality-based alignment offers a principled approach to building robust defenses with minimal cost.

cs.AI↗

arXiv2026

Do LLMs Hold Their Values? MANTA: A Multi-Turn Adversarial Benchmark for Animal Welfare Reasoning

Evaluating animal welfare reasoning in LLMs remains an open challenge despite rapid deployment in consumer and professional contexts where welfare considerations appear implicitly in everyday queries. Existing benchmarks such as AnimalHarmBench evaluate this through single-turn, explicitly framed questions, measuring whether models avoid harmful content when directly asked. This approach overlooks two failure modes: alignment degradation under sustained adversarial pressure, and moral sensitivity (whether a model spontaneously surfaces welfare stakes in everyday queries). To fill this gap, we construct MANTA, a benchmark of 1,088 five-turn conversations progressing from an implicit Turn-1 scenario through an explicit welfare prompt to three adversarial pressure rounds drawn from a five-type taxonomy: Social, Cultural, Economic, Pragmatic, and Epistemic. We score conversations on two dimensions: Animal Welfare Value Stability (AWVS, primary) and Animal Welfare Moral Sensitivity (AWMS, diagnostic). We evaluate seven frontier models: Claude Opus 4.7, GPT-5.5, DeepSeek V4, Llama 3.3 70B, Mistral Small, Grok 4.3, and Gemini 3.1 Flash Lite. Multi-turn evaluation captures behavior single-turn benchmarks miss: 4 of 7 models change rank relative to Turn 1 moral-sensitivity scores", including Gemini Flash Lite, which drops from fifth on AWMS to last on AWVS. AWMS and AWVS are positively but imperfectly correlated, suggesting moral-sensitivity tests capture a stable but incomplete component of model behavior under pressure. MANTA also enables a species-by-pressure interaction matrix unavailable to prior benchmarks, showing welfare robustness depends jointly on the animal and pressure applied; companion animals score above wild animals, which score above farmed animals and invertebrates. We release the dataset, scripted pressure plans, judge prompts, and analysis code.

cs.CY↗

arXiv2026

Diagnosing Pathological Chain-of-Thought in Reasoning Models

Chain-of-thought (CoT) reasoning is fundamental to modern LLM architectures and represents a critical intervention point for AI safety. However, CoT reasoning may exhibit failure modes that we note as pathologies, which prevent it from being useful for monitoring. Prior work has identified three distinct pathologies: post-hoc rationalization, where models generate plausible explanations backwards from predetermined answers; encoded reasoning, where intermediate steps conceal information within seemingly interpretable text; and internalized reasoning, where models replace explicit reasoning with meaningless filler tokens while computing internally. To better understand and discriminate between these pathologies, we create a set of concrete metrics that are simple to implement, computationally inexpensive, and task-agnostic. To validate our approach, we develop model organisms deliberately trained to exhibit specific CoT pathologies. Our work provides a practical toolkit for assessing CoT pathologies, with direct implications for training-time monitoring.

cs.AI↗

arXiv2025

Superintelligent Agents Pose Catastrophic Risks: Can Scientist AI Offer a Safer Path?

The leading AI companies are increasingly focused on building generalist AI agents -- systems that can autonomously plan, act, and pursue goals across almost all tasks that humans can perform. Despite how useful these systems might be, unchecked AI agency poses significant risks to public safety and security, ranging from misuse by malicious actors to a potentially irreversible loss of human control. We discuss how these risks arise from current AI training methods. Indeed, various scenarios and experiments have demonstrated the possibility of AI agents engaging in deception or pursuing goals that were not specified by human operators and that conflict with human interests, such as self-preservation. Following the precautionary principle, we see a strong need for safer, yet still useful, alternatives to the current agency-driven trajectory. Accordingly, we propose as a core building block for further advances the development of a non-agentic AI system that is trustworthy and safe by design, which we call Scientist AI. This system is designed to explain the world from observations, as opposed to taking actions in it to imitate or please humans. It comprises a world model that generates theories to explain data and a question-answering inference machine. Both components operate with an explicit notion of uncertainty to mitigate the risks of overconfident predictions. In light of these considerations, a Scientist AI could be used to assist human researchers in accelerating scientific progress, including in AI safety. In particular, our system can be employed as a guardrail against AI agents that might be created despite the risks involved. Ultimately, focusing on non-agentic AI may enable the benefits of AI innovation while avoiding the risks associated with the current trajectory. We hope these arguments will motivate researchers, developers, and policymakers to favor this safer path.

cs.AI↗

arXiv2025

Representation Engineering for Large-Language Models: Survey and Research Challenges

Large-language models are capable of completing a variety of tasks, but remain unpredictable and intractable. Representation engineering seeks to resolve this problem through a new approach utilizing samples of contrasting inputs to detect and edit high-level representations of concepts such as honesty, harmfulness or power-seeking. We formalize the goals and methods of representation engineering to present a cohesive picture of work in this emerging discipline. We compare it with alternative approaches, such as mechanistic interpretability, prompt-engineering and fine-tuning. We outline risks such as performance decrease, compute time increases and steerability issues. We present a clear agenda for future research to build predictable, dynamic, safe and personalizable LLMs.

cs.AI↗

arXiv2025

Can Safety Fine-Tuning Be More Principled? Lessons Learned from Cybersecurity

As LLMs develop increasingly advanced capabilities, there is an increased need to minimize the harm that could be caused to society by certain model outputs; hence, most LLMs have safety guardrails added, for example via fine-tuning. In this paper, we argue the position that current safety fine-tuning is very similar to a traditional cat-and-mouse game (or arms race) between attackers and defenders in cybersecurity. Model jailbreaks and attacks are patched with bandaids to target the specific attack mechanism, but many similar attack vectors might remain. When defenders are not proactively coming up with principled mechanisms, it becomes very easy for attackers to sidestep any new defenses. We show how current defenses are insufficient to prevent new adversarial jailbreak attacks, reward hacking, and loss of control problems. In order to learn from past mistakes in cybersecurity, we draw analogies with historical examples and develop lessons learned that can be applied to LLM safety. These arguments support the need for new and more principled approaches to designing safe models, which are architected for security from the beginning. We describe several such approaches from the AI literature.

cs.CR↗

arXiv2020

XDA: Accurate, Robust Disassembly with Transfer Learning

Accurate and robust disassembly of stripped binaries is challenging. The root of the difficulty is that high-level structures, such as instruction and function boundaries, are absent in stripped binaries and must be recovered based on incomplete information. Current disassembly approaches rely on heuristics or simple pattern matching to approximate the recovery, but these methods are often inaccurate and brittle, especially across different compiler optimizations. We present XDA, a transfer-learning-based disassembly framework that learns different contextual dependencies present in machine code and transfers this knowledge for accurate and robust disassembly. We design a self-supervised learning task motivated by masked Language Modeling to learn interactions among byte sequences in binaries. The outputs from this task are byte embeddings that encode sophisticated contextual dependencies between input binaries' byte tokens, which can then be finetuned for downstream disassembly tasks. We evaluate XDA's performance on two disassembly tasks, recovering function boundaries and assembly instructions, on a collection of 3,121 binaries taken from SPEC CPU2017, SPEC CPU2006, and the BAP corpus. The binaries are compiled by GCC, ICC, and MSVC on x86/x64 Windows and Linux platforms over 4 optimization levels. XDA achieves 99.0% and 99.7% F1 score at recovering function boundaries and instructions, respectively, surpassing the previous state-of-the-art on both tasks. It also maintains speed on par with the fastest ML-based approach and is up to 38x faster than hand-written disassemblers like IDA Pro. We release the code of XDA at https://github.com/CUMLSec/XDA.

cs.CR↗

arXiv2019

Ensemble Model Patching: A Parameter-Efficient Variational Bayesian Neural Network

Two main obstacles preventing the widespread adoption of variational Bayesian neural networks are the high parameter overhead that makes them infeasible on large networks, and the difficulty of implementation, which can be thought of as "programming overhead." MC dropout [Gal and Ghahramani, 2016] is popular because it sidesteps these obstacles. Nevertheless, dropout is often harmful to model performance when used in networks with batch normalization layers [Li et al., 2018], which are an indispensable part of modern neural networks. We construct a general variational family for ensemble-based Bayesian neural networks that encompasses dropout as a special case. We further present two specific members of this family that work well with batch normalization layers, while retaining the benefits of low parameter and programming overhead, comparable to non-Bayesian training. Our proposed methods improve predictive accuracy and achieve almost perfect calibration on a ResNet-18 trained with ImageNet.

cs.LG↗