Research map / Large language models / Reasoning and chain of thought

The Art of Interrogation: Consistency Amplifies Factuality in Spatial Reasoning

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 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 label-free consistency-verifier RL (OT-GRPO, minimal-matching GRPO over paired completions from original and transformed prompts) that nearly matches ground-truth accuracy RL for spatial reasoning in Qwen2.5-VL.

Problem

Large reasoning and vision-language models underperform on spatial reasoning, and existing fixes rely on labeled spatial data from vision pipelines or synthetic engines; the paper asks whether the capability can be unlocked without ground-truth labels.

Approach

Use known transformation-induced relations between answers (invariance or negation under flips, crops, object swaps, relation swaps) as a self-supervised reward, and pair completions from original and augmented prompts via an optimal-transport minimal-consistency matching inside GRPO (OT-GRPO).

Key results

• Consistency-trained Qwen2.5-VL-7B reaches 76.8% vs 79.6% for accuracy training on SUN RGB-D same-task evaluation (gap 2.8pp), baseline 63.4%.
• On joint four-task training, consistency reward reaches 83.8% vs 86.5% for accuracy training and beats Visual Jigsaw and SSL4RL checkpoints by +3.6pp and +13-24pp.
• Consistency overtakes accuracy training under 20%+ uniform label corruption (e.g., 76.4% vs 68.8%-equivalent gap of 7.4pp at 40% noise per paper's reported gaps).

Surprising findings

• Label-free consistency reward overtakes accuracy training once 20% or more of training labels are randomly flipped.
• Minimal (OT) pairing yields a random-guesser expected reward of about 1/sqrt(pi K), penalizing uninformative models unlike random or one-to-all pairing.

Limitations

• Consistency rewards compare only two predictions at a time; chains of three or more transformations are not explored.
• Experiments are limited to binary spatial tasks plus two numeric tasks, and the consistency-trained model still trails accuracy training by about 2.7pp.
• Label-corruption noise model is uniform, which the authors note is conservative relative to correlated real annotation errors.

Experimental setup

Datasets: SUN RGB-D KITTI Omni3D

Benchmarks: SUN RGB-D held-out split (orientation, depth, size, relative distance) KITTI held-out split (cross-domain) counting (SUN RGB-D) absolute distance (SUN RGB-D)

Metrics: accuracy improvement over baseline (pp) numeric rescaled MAE-based score

Baselines: Visual Jigsaw SSL4RL accuracy training (ground-truth RL) random pairing one-to-all pairing

Builds on: GRPO DeepSeek-R1 Omni3D LoRA

Closest papers by meaning

• Active Exploring like a Pigeon: Reinforcing Spatial Reasoning via Agentic Vision-Language Models (ICML 2026 · similarity 0.88)
• Spatial-SSRL: Enhancing Spatial Understanding via Self-Supervised Reinforcement Learning (CVPR 2026 · similarity 0.88)
• Fine-Grained Preference Optimization Improves Spatial Reasoning in VLMs (NeurIPS 2025 · similarity 0.88)
• SpatialLadder: Progressive Training for Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.87)
• Geometrically-Constrained Agent for Spatial Reasoning (CVPR 2026 · similarity 0.87)
• No Labels, No Problem: Training Visual Reasoners with Multimodal Verifiers (ICLR 2026 · similarity 0.87)