Research map / Vision-language and multimodal / Multimodal LLMs

No Labels, No Problem: Training Visual Reasoners with Multimodal Verifiers

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes VALOR, an annotation-free framework that trains an LLM for tool-using visual reasoning via GRPO with LLM-verifier rewards and fine-tunes GroundingDINO via VLM-verified pseudo-labels from hard-negative over-prediction.

Problem

Visual reasoning requires precise grounding and complex spatial logic, but existing VLM chain-of-thought methods need large labeled supervision and program-synthesis methods rely on proprietary LLMs and poorly aligned pre-trained vision specialists.

Approach

Use AI verifiers in place of ground truth: a frozen LLM verifier provides structured multi-part rewards (format, syntax, logic, attribute, spatial, adherence) to train a Qwen3-8B planner-coder with GRPO, while a frozen VLM verifier filters over-predicted GroundingDINO boxes in three stages (coarse filter, per-crop check, de-duplication) to create pseudo-labels for fine-tuning the detector.

Key results

• VALOR on Omni3D-Bench reaches 44.0% vs 37.5% for the base Qwen3-8B and 38.9% for VADAR, and outperforms GRIT (27.3%) and ViGoRL (13.2%).
• VALOR reaches 69.5% on RoboSpatial vs 60.9% for Gemini-2.0-Flash and 64.9% for ViGoRL, and 75.9% on CountBenchQA, trailing GPT-4o (84.7%) and Gemini-2.0-Flash (88.6%).
• Gemini-2.5-Flash as LLM verifier agrees with manual annotations on 87% of 100 outputs, versus 15% for Qwen3-8B and 7% for Llama-3.2-11B-Instruct.

Surprising findings

• Open-source Qwen3-8B with tools and verifier-trained grounding outperforms proprietary GPT-5-mini on Omni3D-Bench (44.0% vs 40.9%).
• GPT-5-mini as a verifier is strong at judging detections but its own grounded bounding boxes are poor, leading to lower Omni3D-Bench accuracy (35.5%) when grounded.

Limitations

• Verifier-based pseudo-labels reach only 75% precision and the detector struggles with overlapping similar objects.
• Depends on proprietary Gemini-2.5-Flash and GPT-5-mini as verifiers; CountBenchQA trails direct-prediction VLMs.
• Reasoning failures persist on challenging hypothetical queries in Omni3D-Bench.

Experimental setup

Datasets: Omni3D-Bench TallyQA GQA VSR SA-1B

Benchmarks: Omni3D-Bench RoboSpatial BLINK VSR RealWorldQA GQA TallyQA CountBenchQA COCO

Metrics: exact-match accuracy MRA mAP

Baselines: GPT-4o o4-mini Gemini-2.0-Flash Gemini-2.5-Flash Claude-3.5-Haiku Llama-3.2-11B Gemma-3-12B Qwen3-8B GRIT ViGoRL VisProg ViperGPT

Builds on: GRPO GroundingDINO VisProg ViperGPT VADAR

Closest papers by meaning

• Connecting the Dots: Training-Free Visual Grounding via Agentic Reasoning (AAAI 2026 · similarity 0.91)
• VLM-R³: Region Recognition, Reasoning, and Refinement for Enhanced Multimodal Chain-of-Thought (NeurIPS 2025 · similarity 0.91)
• Multi-step Visual Reasoning with Visual Tokens Scaling and Verification (NeurIPS 2025 · similarity 0.91)
• VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.91)
• RegionReasoner: Region-Grounded Multi-Round Visual Reasoning (ICLR 2026 · similarity 0.91)
• Reinforcing Spatial Reasoning in Vision-Language Models with Interwoven Thinking and Visual Drawing (NeurIPS 2025 · similarity 0.91)