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

Self-Consistency for LLM-Based Motion Trajectory Generation and Verification

CVPR 2026 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling · 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.

Adapts self-consistency to LLM-generated motion graphics trajectories by clustering samples under a hierarchy of Lie transformation groups, selecting the group and prototype that best explain the largest cluster, and using membership tests for verification.

Problem

LLM-generated motion trajectories for a text prompt are hard to compare with identity matching, since a prompt describes a family of shapes rather than one trajectory; this blocks unsupervised self-consistency for visual generation and verification.

Approach

Model each prompt's shape family as a prototype trajectory plus a Lie transformation group; define warp-invariant distances per group and cluster LLM samples with DBSCAN. Two decision criteria (Majority-Consensus and Hierarchical-Consistency) estimate the group from a hierarchy of groups, and membership in the largest cluster is used for verification.

Key results

• Self-consistency with Majority-Consensus raises generation accuracy from 62.1% to 68.0% for GPT-4.1 and from 79.1% to 83.3% for GPT-5.
• Hierarchical-Consistency achieves the best verification F1 of 84.6 versus 79.0 for GPT-5 as a VLM verifier.
• Majority-Consensus reaches 85.8 precision versus 74.0 for GPT-5 verifier, an 11.8-point gain.

Surprising findings

• Majority-Consensus tends toward overly restrictive groups and Hierarchical-Consistency toward overly permissive ones when wrong.
• For 16.1% (GPT-5) and 31.3% (GPT-4.1) of prompts, the largest cluster contains no correct trajectories even under oracle warps.

Limitations

• Cannot recover from sampling-LLM mistakes when incorrect samples form the largest cluster.
• Fails for ambiguous prompts or multi-prototype families such as heptagrams without extension.
• Shape families with anisotropic similarity require canonical orientation samples from the LLM.

Experimental setup

Datasets: template-based motion trajectory benchmark (224 prompts, 35 base shapes)

Benchmarks: motion trajectory benchmark (224 prompts) motion trajectory verification test set (2240 trajectories)

Metrics: accuracy precision recall F1

Baselines: LLM-Direct GPT-4.1 verifier GPT-5 verifier Most-Restrictive Least-Restrictive

Builds on: self-consistency DBSCAN iterative closest point MoVer

Closest papers by meaning

• ConsisVLA-4D: Advancing Spatiotemporal Consistency in Efficient 3D-Perception and 4D-Reasoning for Robotic Manipulation (CVPR 2026 · similarity 0.85)
• Controllable Dynamic 3D Shape Generation via 3D Trajectories and Text (NeurIPS 2026 · similarity 0.85)
• Can Large Language Models Understand Symbolic Graphics Programs? (ICLR 2025 · similarity 0.84)
• Riemannian Consistency Model (NeurIPS 2025 · similarity 0.84)
• AvatarGPT: All-in-One Framework for Motion Understanding Planning Generation and Beyond (CVPR 2024 · similarity 0.84)
• MolmoMotion: Forecasting Point Trajectories in 3D with Language Instruction (NeurIPS 2026 · similarity 0.83)