arXiv subjects

Ali Asad

Publications and source records attributed to Ali Asad.

3 recordsLinked to original sources

arXiv2026

Confidently Deceptive: How Confidence Amplifies the Risk of LLM Deception

Large language models (LLMs) can produce deceptive responses: outputs that mislead users in service of a contextually or experimentally induced goal. Yet it remains unclear how confidently models deceive and whether higher confidence makes deceptive responses more persuasive to end users. In this paper, we study these basic questions in various models and different deception datasets. We provide a comprehensive study measuring confidence through both verbalized self-reports and a range of logit-based estimators. We show that LLMs deliver deceptive responses with substantial verbalized confidence and that human annotators prefer the higher-confidence deceptive response 78% of the time in paired comparisons. Misalignment fine-tuning amplifies the problem. Confidence in deceptive responses rises across all three benchmarks, increasing the resulting potential risk, with effects generalizing beyond the training distribution. Strikingly, models classify their own deceptive outputs as deceptive at high rates (82.7% under misalignment) while still predicting they would produce them - recognition without avoidance. We argue that confident deception is a distinct alignment risk requiring evaluations that jointly measure deception, confidence, and awareness.

cs.CL↗

arXiv2025

RedDebate: Safer Responses Through Multi-Agent Red Teaming Debates

We introduce RedDebate, a novel multi-agent debate framework that provides the foundation for Large Language Models (LLMs) to identify and mitigate their unsafe behaviours. AI safety approaches often rely on costly human evaluation or isolated single-model assessment, both constrained by scalability and prone to oversight failures. RedDebate employs collaborative argumentation among multiple LLMs across diverse debate scenarios, enabling them to critically evaluate one another's reasoning and systematically uncover unsafe failure modes through fully automated red-teaming. To support this, we propose designing distinct long-term memory modules that preserve safety-relevant insights from debate interactions and leverage them during subsequent inference, facilitating continuous refinement of model behaviour. Empirical evaluation on safety benchmarks across a diverse set of models demonstrates that RedDebate substantially reduces unsafe outputs. While debate alone allows LLMs to refine their behaviour, the addition of memory yields further error reductions. To the best of our knowledge, RedDebate is the first fully automated framework to unify multi-agent debate and red-teaming to progressively enhance LLM safety without human intervention.

cs.CL↗

arXiv2023

High-order adaptive multi-domain time integration scheme for microscale lithium-ion batteries simulations

We investigate the modeling and simulation of ionic transport and charge conservation in lithium-ion batteries (LIBs) at the microscale. It is a multiphysics problem that involves a wide range of time scales. The associated computational challenges motivate the investigation of numerical techniques that can decouple the time integration of the governing equations in the liquid electrolyte and the solid phase (active materials and current collectors). First, it is shown that semi-discretization in space of the non-dimensionalized governing equations leads to a system of index-1 semi-explicit differential algebraic equations (DAEs). Then, a new generation of strategies for multi-domain integration is presented, enabling high-order adaptive coupling of both domains in time, with efficient and potentially different domain integrators. They reach a high level of flexibility for real applications, beyond the limitations of multirate methods. A simple 1D LIB half-cell code is implemented as a demonstrator of the new strategy for the simulation of different modes of cell operation. The integration of the decoupled subsystems is performed with high-order accurate implicit nonlinear solvers. The accuracy of the space discretization is assessed by comparing the numerical results to the analytical solutions. Then, temporal convergence studies demonstrate the accuracy of the new multi-domain coupling approach. Finally, the accuracy and computational efficiency of the adaptive coupling strategy are discussed in the light of the conditioning of the decoupled subproblems compared to the one of the fully-coupled problem. This new approach will constitute a key ingredient for the high-fidelity 3D LIB simulations based on actual electrode microstructures.

math.NA↗