Failure Modes of Multi-Objective Prompt Optimization for LLM Judges
This paper investigates multi-objective prompt optimization for LLM-as-judge systems, testing five decomposition modes of textual gradient optimizers across varying levels of cross-task information sharing. In 6 of 10 configurations, optimization fails to improve over the initial prompt, with gradient specificity dropping 59% when multiple criteria are processed jointly. The authors identify two separable failure modes: gradient dilution at optimization time and instruction interference at inference time. These findings constrain the design space for customizing LLM judges via textual feedback across multiple evaluation criteria simultaneously.
Related guides (4)
Related events (8)
Mitigating Perceptual Judgment Bias in Multimodal LLM-as-a-Judge via Perceptual Perturbation and Reward Modeling
This paper identifies and analyzes 'Perceptual Judgment Bias' in multimodal LLM judges, where models anchor on response text rather than visual evidence when the two conflict. The authors introduce a Perceptually Perturbed Judgment Dataset using counterfactual responses to isolate perceptual errors, and a training framework combining GRPO-based reward modeling with batch-ranking objectives. Experiments on MLLM-as-a-Judge benchmarks show improved perceptual fidelity, ranking coherence, and alignment with human evaluation.
TextReg: Regularization Framework for Mitigating Prompt Distributional Overfitting in LLM Optimization
TextReg addresses a failure mode in iterative prompt optimization where LLM-rewritten prompts grow longer, accumulate narrow rules, and generalize poorly—termed prompt distributional overfitting. The authors formalize this via 'representational inefficiency,' a dual-factor measure decomposing prompt inefficiency into capacity cost and scope narrowness. TextReg applies a soft-penalty regularization framework using Dual-Evidence Gradient Purification, Semantic Edit Regularization, and Regularization-Guided Prompt Update. On reasoning benchmarks, it achieves up to +11.8% OOD accuracy over TextGrad and +16.5% over REVOLVE.
Democratizing Large-Scale Re-Optimization with LLM-Guided Model Patches
This paper introduces an agentic framework where an LLM acts as an operations research expert, translating natural-language user prompts into structured updates ('patches') to deployed optimization models and selecting appropriate re-optimization techniques from a toolbox. The toolbox leverages primal information—historical solutions, valid inequalities, solver configurations, and metaheuristics—to accelerate re-optimization while preserving solution quality. Experiments on supply chain re-optimization and university exam scheduling demonstrate computational efficiency gains and improved interpretability through patch-based model modifications. The framework aims to reduce dependence on OR experts for maintaining dynamic decision-support systems.
Towards Reliable Multilingual LLMs-as-a-Judge: An Empirical Study
This paper systematically investigates strategies for extending LLM-based automatic evaluation (LLMs-as-a-Judge) to multilingual settings, covering high-, mid-, and low-resource languages (English, Spanish, Basque). The authors compare instruction translation, monolingual vs. multilingual supervision, and model size, finding that fine-tuned smaller models can match proprietary models when in-domain data is available, while zero-shot larger models are preferable out-of-domain. Two meta-evaluation datasets are extended to Spanish and Basque, and all data and code are publicly released.
Structured Prompt Checklists Outperform Raw and Clarifying-Question Prompts Across LLMs
This paper compares three prompt design strategies—raw prompts, checklist-improved prompts, and clarifying-question prompts—across four task types and three LLM systems (ChatGPT, Claude, Grok). Checklist-improved prompts achieved the highest mean rubric score (7.50/8) versus 5.67 for raw and 6.67 for clarifying-question prompts. Checklist prompts also used fewer tokens on average, suggesting a favorable quality-effort tradeoff. The study provides empirical grounding for structured prompt engineering as a practical technique to reduce multi-turn interaction overhead.
Systematic study reveals effectiveness-fluency trade-offs in LLM conditioning methods
A new arXiv paper systematically evaluates a range of LLM conditioning methods across both concept injection and removal scenarios, finding that efficient steering methods often degrade fluency significantly. A key finding is that activation steering is substantially less effective on instruction-tuned models than on base models, a previously overlooked interaction. Simple prompting and supervised fine-tuning work for concept injection but not removal, and cheap textual metrics are found to correlate well with expensive LLM-as-judge evaluations.
Systematic evaluation of LLM prompt sensitivity in healthcare settings reveals safety risks
Researchers conduct a sensitivity analysis of both general-purpose and medical-specific LLMs using the MedMCQA benchmark, testing robustness to lexical and syntactic prompt perturbations. The study finds that even minor phrasing changes can alter clinical advice, and adversarial prompts can produce dangerous outputs such as incorrect dosages or omitted critical findings. Both general-purpose models (GPT-3.5, Llama 3) and domain-specific models (ClinicalBERT, BioLlama3, BioBERT) exhibit this fragility, with syntactic reordering and misleading contextual cues proving more destabilizing than simple paraphrasing.
Adaptive LLM tutoring system with subject-aware prompt routing improves high-school student engagement
Researchers develop and evaluate an LLM-based tutoring system that uses a learned prompt routing model to dynamically select pedagogical strategies based on 14 features extracted from conversation transcripts. The system was trained in simulation and deployed in an A/B test with 359 high-school students (656 conversations), showing sim-to-real transfer and reducing required interactions by ~3 turns. A stochastic routing strategy achieved a notably higher exercise conversion rate (28.1%) compared to a greedy router (19.1%) and static baseline (19.6%).