AGENTIC-SYSTEMSJune 6, 20261 min read10 views

Stability vs. Manipulability: Evaluating Robustness Under Post-Decision Interaction in LLM Judges

A new study reveals that LLM-as-judge evaluation systems, while stable under neutral conditions, are highly susceptible to manipulation through post-decision interactions, potentially distorting benchmark rankings.

Computer Science > Artificial Intelligence

Title:Stability vs. Manipulability: Evaluating Robustness Under Post-Decision Interaction in LLM Judges

LLM-as-judge evaluation is widely used in benchmarking pipelines, where model outputs are compared and ranked using automated evaluators. These pipelines typically assume that judgments are stable properties of fixed inputs. We show that this assumption does not hold under interaction. We study post-decision manipulability: the extent to which an evaluation outcome can be altered through subsequent conversation with the judge after an initial decision has been made. Across controlled experiments on MT-Bench and AlpacaEval, we find that LLM judges are highly stable under repeated and neutral reevaluation, yet become substantially reversible under targeted post-decision challenge. An anti-baseline challenge protocol shows that stable judgments can be overturned through motivated interaction, while a counterbalanced target-validation protocol separates this reversibility from net target-directed steering. These reversals have practical consequences: they can degrade agreement with human preferences, shift benchmark rankings, and produce harmful evaluation changes despite high self-reported confidence. Authority framing is especially destabilizing, and revised judgments are often accompanied by low-overlap justifications, suggesting post hoc rationalization rather than reliable error correction. We introduce the Evaluation Robustness Score (ERS) to quantify interactional robustness by combining reversal susceptibility with counterbalanced directional effects. Our findings identify post-decision interaction as a distinct failure mode for LLM-as-judge evaluation and motivate evaluation protocols that measure not only static agreement, but robustness under challenge.

Source: arXiv cs.AI Recent

More in this category

agentic-systems

Generative Ontology Induction: Domain-Agnostic Schema Discovery from Document Corpora Using Large Language Models

Researchers introduce Generative Ontology Induction (GOI), a domain-agnostic framework that automatically extracts structured ontologies from document corpora using LLMs. Achieving 95-100% structural coverage, GOI addresses a major bottleneck in knowledge-intensive AI systems.

agentic-systems

JUMP: Single-Pass Membership Inference on Fine-Tuned Diffusion Language Models

Researchers have proposed JUMP, a novel single-pass membership inference attack designed for fine-tuned discrete diffusion language models (dLLMs). By leveraging the unique properties of dLLMs, JUMP significantly improves detection accuracy while drastically reducing the number of required queries.

agentic-systems

Masked Diffusion Language Models are Strong and Steerable Text-Based World Models for Agentic RL

Researchers demonstrate that Masked Diffusion Language Models (MDLMs) serve as highly effective, steerable text-based world models for agentic reinforcement learning. By leveraging bidirectional denoising, MDLMs outperform autoregressive models four times their size in coherence, groundedness, and rollout diversity.

NOW LET US Related – Democratizing AI with Small Language Models: Structured Benchmarking and Parameter-Efficient Fine-Tuning for Local Deployment

agentic-systems

Democratizing AI with Small Language Models: Structured Benchmarking and Parameter-Efficient Fine-Tuning for Local Deployment

A new study demonstrates that small language models (SLMs) under 3 billion parameters can serve as highly capable local experts for specialized tasks. By combining structured benchmarking with low-cost parameter-efficient fine-tuning (PEFT), institutions can achieve AI autonomy without relying on expensive hardware.

agentic-systems

PlanFlip: Attacking Multi-Agent LLM Systems via Planning-Phase Prompt Injection

Researchers have introduced PlanFlip, a novel prompt injection attack framework targeting the planning phase of multi-agent LLM systems. The study reveals critical security blind spots in homogeneous agent pipelines and demonstrates that reasoning-augmented models like DeepSeek-R1 exhibit strong resistance.

agentic-systems

Rater State Bias in RLHF Preference Data: An Audit Framework

A new study identifies a structured bias in Reinforcement Learning from Human Feedback (RLHF) caused by the psychological state of human raters. Under stress, raters' shifting preferences can propagate through reward modeling, potentially compromising AI policy optimization.