Let's Think in Two Steps: Mitigating Agreement Bias in MLLMs with Self-Grounded Verification
cs.AI, cs.CL, cs.LG, cs.MA, cs.RO
Submitted: 2025-07-15
Updated: 2026-11-11
Comments: ICLR 2026. Code, models, and data publicly available at https://mshalimay.github.io/agreement-bias-sgv/
Journal ref: Proceedings of the Fourteenth International Conference on Learning Representations (ICLR 2026)
Code: https://github.com/kaistAI/prometheus-vision
Project page: https://mllm-judge.github.io
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Terminology
Sources
- Physics of Language Models: Part 3.1, Knowledge Storage and Extraction
- Physics of Language Models: Part 3.2, Knowledge Manipulation
- DigiRL: Training In-The-Wild Device-Control Agents with Autonomous Reinforcement Learning
- Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation
- Do LLM Evaluators Prefer Themselves for a Reason?
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
- Large Language Models Cannot Self-Correct Reasoning Yet
- LLMs Can't Plan, But Can Help Planning in LLM-Modulo Frameworks
- Kimi k1.5: Scaling Reinforcement Learning with LLMs
- VisualWebArena: Evaluating Multimodal Agents on Realistic Visual Web Tasks
- Tree Search for Language Model Agents
- Prometheus-Vision: Vision-Language Model as a Judge for Fine-Grained Evaluation
- BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models
- On the Effects of Data Scale on UI Control Agents
- Let's Verify Step by Step
- AgentBench: Evaluating LLMs as Agents
- Vision Language Models are In-Context Value Learners
- What Matters in Learning from Offline Human Demonstrations for Robot Manipulation
- Agent-RewardBench: Towards a Unified Benchmark for Reward Modeling across Perception, Planning, and Safety in Real-World Multimodal Agents
- Autonomous Evaluation and Refinement of Digital Agents
Related papers
- Multi-Objective Structured Pruning of LLMs for Latency and Model Size Optimization
- MAVEN-T: Reinforced Heterogeneous Distillation for Real-Time Multi-Agent Trajectory Prediction
- Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models
- Embeddings for Preferences, Not Semantics
- Can We Trust Item Response Theory for AI Evaluation?
- Diagnosing Task Insensitivity in Language Agents