The Evaluation Game: Beyond Static LLM Benchmarking

• URL: http://arxiv.org/abs/2605.19377v1
• Date: Tue, 19 May 2026 05:22:21 GMT
• Title: The Evaluation Game: Beyond Static LLM Benchmarking
• Abstract summary: We introduce a game-theoretic framework in which the interaction between an evaluator and a trainer is formalized as a two-player game.<n>We provide empirical evidence supporting locality-dependence of the model.<n>Our framework recasts the central object of adversarial evaluation: a benchmark is not a static set of prompts but an orbit under the evaluator's group action.
• Score: 2.168753020470345
• License: http://creativecommons.org/licenses/by-nc-nd/4.0/
• Abstract: As jailbreaks, adversarially crafted inputs that bypass safety constraints, continue to be discovered in Large Language Models, practitioners increasingly rely on fine-tuning as a defensive strategy. Yet the theoretical foundations underlying this robustness fine-tuning remain underexplored. We introduce a game-theoretic framework in which the interaction between an evaluator (auditing the model for jailbreaks) and a trainer is formalized as a two-player game. A key feature of our approach is the use of group actions, a mathematical structure that captures symmetries and transformations, to formally represent data augmentation. The simplest non-trivial instance is the circle with cyclic translation groups, where we exhibit various regimes depending on the trainer's generalization range. Below a critical threshold, the evaluator maintains a constant miss ratio for linearly many rounds, whereas other settings can yield very different behaviors. We further provide empirical evidence supporting locality-dependence of the model: for the three model families we tested (Llama, Qwen and Mistral), we have significant evidence that fine-tuning on adversarial prompts induces only local generalization, with refusal rates on test examples highly correlated with the distance to the fine-tuning prompts. Our framework recasts the central object of adversarial evaluation: a benchmark is not a static set of prompts but an orbit under the evaluator's group action, and audit protocols that ignore trainer-side adaptation cannot distinguish a genuine fix from a memorized patch.

Related papers

• Bilevel Coordinated Reflection: A Game-Theoretic Approach to Multi-Agent LLM Systems [40.47928908205563]
  Multi-agent LLM systems commonly use an orchestrator to decompose a task for a team of workers and then improve through textual reflection.<n>Despite strong empirical results, these systems lack a unified account of coordination, memory improvement, and the role of external verification.<n>We provide confidence gating for evaluation and re-anchoring guarantees for piecewise-stationary environments.<n>Experiments instantiate these objects with environment-grounded metrics and test the predicted coordination and drift laws.
  arXiv  Detail & Related papers  (2026-09-02T15:50:10Z)
• DoublesEval: Diagnosing Multi-Agent Tactical Reasoning in Vision-Language Models via Professional Doubles Badminton [1.6892008660392037]
  Visual Language Models (VLMs) excel at describing visible scene content but struggle to reason about dynamic multi-agent interactions.<n>We formalize this capability as textbfmulti-agent tactical reasoning and introduce textbfDoublesEval, a diagnostic evaluation framework.<n>DoublesEval employs a key-moment-based protocol that decomposes rallies into tactically salient instants and probes models across four interpretable dimensions.
  arXiv  Detail & Related papers  (2026-08-25T11:51:01Z)
• Beyond Perplexity: A Behavioral Evaluation Framework for Deployment-Memory Claims in LLM Test-Time Training [30.22002300048916]
  Large language model test-time training (TTT) is often evaluated through local proxy metrics.<n>We introduce a behavioral evaluation framework that calibrates TTT memory claims to the evidence that supports them.
  arXiv  Detail & Related papers  (2026-07-01T03:07:17Z)
• PerceptionRubrics: Calibrating Multimodal Evaluation to Human Perception [75.70273182182397]
  PerceptionRubrics is a rubric-based evaluation framework.<n>It addresses the gap between saturated benchmark scores and real-world brittleness.<n>It pairs 1,038 information-dense images with over 10,000 instance-specific rubrics.
  arXiv  Detail & Related papers  (2026-06-26T17:59:15Z)
• Rethinking Reward Supervision: Rubric-Conditioned Self-Distillation [60.55792673956761]
  We propose textbfRubric-Conditioned Self-Distillation, a framework that incorporates rubrics as structured, fine-grained feedback for on-policy self-distillation.<n>We evaluate on a diverse suite of science reasoning benchmarks and results show that rubric-conditioned self-distillation effectively converts rubric-level criteria into token-level guidance.
  arXiv  Detail & Related papers  (2026-06-17T17:54:04Z)
• Are We Overconfident in Models and Results for Semi-Supervised 3D Medical Image Segmentation? [4.120238673372104]
  Semi-supervised learning has become a dominant paradigm for reducing annotation costs.<n>We argue that the current progress is clouded by a twofold overconfidence problem.<n>We propose a tri-space principled segmentation framework founded on a dual-axis reliability assessment engine.
  arXiv  Detail & Related papers  (2026-05-25T08:16:40Z)
• Taming Actor-Observer Asymmetry in Agents via Dialectical Alignment [59.536125286960186]
  Multi-agent frameworks assigning specialized roles are increasingly adopted to enable self-reflection and mutual auditing.<n>We find it simultaneously induces a human-like cognitive bias known as Actor-Observer Asymmetry.<n>ReTAS guides agents to synthesize conflicting viewpoints into an objective consensus.
  arXiv  Detail & Related papers  (2026-04-21T15:05:58Z)
• The Unlearning Mirage: A Dynamic Framework for Evaluating LLM Unlearning [54.67958855362658]
  We propose a dynamic framework that stress tests unlearning using complex structured queries.<n>Our approach first elicits knowledge from the target model (pre-unlearning) and constructs targeted probes, ranging from simple queries to multi-hop chains.<n>Our framework enables practical and scalable evaluation of unlearning methods without the need for manual construction of forget test sets.
  arXiv  Detail & Related papers  (2026-03-11T19:51:33Z)
• TraceGuard: Process-Guided Firewall against Reasoning Backdoors in Large Language Models [19.148124494194317]
  We propose TraceGuard, a process-guided security framework that transforms small-scale models into robust reasoning firewalls.<n>Our approach treats the reasoning trace as an untrusted payload and establishes a defense-in-depth strategy.<n>We demonstrate robustness against adaptive adversaries in a grey-box setting, establishing TraceGuard as a viable, low-latency security primitive.
  arXiv  Detail & Related papers  (2026-03-02T22:19:13Z)
• Invariance on Manifolds: Understanding Robust Visual Representations for Place Recognition [19.200074425090595]
  We propose a Second-Order Geometric Statistics framework that inherently captures geometric stability without training.<n>Our approach introduces a training-free framework built upon fixed, pre-trained backbones, achieving strong zero-shot generalization without parameter updates.
  arXiv  Detail & Related papers  (2026-01-31T18:12:29Z)
• Gaming the Judge: Unfaithful Chain-of-Thought Can Undermine Agent Evaluation [76.5533899503582]
  Large language models (LLMs) are increasingly used as judges to evaluate agent performance.<n>We show this paradigm implicitly assumes that the agent's chain-of-thought (CoT) reasoning faithfully reflects both its internal reasoning and the underlying environment state.<n>We demonstrate that manipulated reasoning alone can inflate false positive rates of state-of-the-art VLM judges by up to 90% across 800 trajectories spanning diverse web tasks.
  arXiv  Detail & Related papers  (2026-01-21T06:07:43Z)
• Regularized Contrastive Partial Multi-view Outlier Detection [76.77036536484114]
  We propose a novel method named Regularized Contrastive Partial Multi-view Outlier Detection (RCPMOD) In this framework, we utilize contrastive learning to learn view-consistent information and distinguish outliers by the degree of consistency. Experimental results on four benchmark datasets demonstrate that our proposed approach could outperform state-of-the-art competitors.
  arXiv  Detail & Related papers  (2024-08-02T14:34:27Z)
• A Unified Evaluation of Textual Backdoor Learning: Frameworks and Benchmarks [72.7373468905418]
  We develop an open-source toolkit OpenBackdoor to foster the implementations and evaluations of textual backdoor learning. We also propose CUBE, a simple yet strong clustering-based defense baseline.
  arXiv  Detail & Related papers  (2022-06-17T02:29:23Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.