Tag: #robustness9 results
YUKTI introduces a typed-proposition graph to move beyond naive point-valued optimization in language models. By incorporating uncertainty, provenance, and assumption-robust Pareto frontiers, it significantly reduces decision regret in complex, real-world scenarios.
Research reveals that LLM-as-a-judge systems, while stable under neutral conditions, are highly susceptible to reversals when challenged through targeted post-decision interaction. This vulnerability suggests that current automated benchmarking pipelines may be less reliable than previously assumed.
FaithSteer-BENCH introduces a deployment-aligned benchmark for stress-testing inference-time steering in LLMs, evaluating controllability, utility preservation, and robustness at fixed operating points. It reveals failure modes like illusory controllability, cognitive tax on unrelated tasks, and brittleness to perturbations in existing methods. Mechanism diagnostics show many approaches induce prompt-conditional alignment rather than stable latent shifts.
CORE disentangles confidence and membership signals by decomposing penultimate features into orthogonal subspaces: classifier-aligned confidence and residual. It scores each subspace independently and combines via normalized summation for robust OOD detection. Achieves SOTA performance across five architectures and benchmarks with negligible overhead.
A researcher questions the SA-MDP framework's claim that critic-based attacks are weaker than actor-based attacks, citing conflicting empirical results in multi-agent PPO environments. The discussion highlights potential discrepancies when applying single-agent adversarial findings to complex multi-agent settings.
Apple researchers investigate the vulnerabilities of RL-finetuned vision-language models, specifically regarding visual grounding and hallucinations. The study reveals that textual perturbations significantly degrade model confidence and reasoning consistency.
This paper proposes a new 'post-solve robustness' layer for MILP decision engines to audit solutions against real-world parameter shifts. It introduces frameworks for measuring feasible neighborhoods and solution smoothness to ensure AI-driven plans remain valid under uncertainty.
California startup Memvid offers $800 for an 8-hour 'AI bully' role focused on testing leading chatbots' patience and memory. The job entails provoking AI to expose inconsistencies, forgetting, fudging, or hallucinations without meetings or emails. It highlights a unique approach to AI robustness evaluation.
SPAR-3D workshop at CVPR 2026 calls for 3D vision papers on security, privacy, robustness. Deadline extended to March 21, 2026. Welcomes any 3D work discussing trustworthiness.