Research map / Vision-language and multimodal / Multimodal LLMs

STELAR-VISION: Self-Topology-Aware Efficient Learning for Aligned Reasoning in Vision

AAAI 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 STELAR-Vision, which generates chain, tree and graph reasoning responses (TopoAug) for math VQA and post-trains Qwen2-VL-7B with SFT, SimPO-based RL, and a length-reducing Frugal Learning variant.

Problem

VLMs default to verbose chain-of-thought reasoning even though different questions benefit from different reasoning topologies (chain, tree, graph), and they overthink, increasing cost.

Approach

TopoAug synthesizes multiple responses per question under chain, tree, and graph prompts, labels per-question topology accuracy and per-response correctness, and trains the VLM (SFT then SimPO preference RL) to select a suitable topology implicitly at test time; Frugal Learning adds length-aware preference pairs to shorten outputs.

Key results

• On in-distribution VLM_S2H and MATH-V, STELAR-Vision reaches 28.0% overall accuracy versus 18.3% for the Qwen2VL-7B-Instruct base model (+9.7 points) and 20.7% for Qwen2VL-72B-Instruct.
• Compared with Chain-Only training, STELAR-Vision gains 4.3 points overall on the ID datasets (23.7% to 28.0%).
• Frugal Learning variant STELAR-Vision-Short† cuts ID generated tokens from 556.7 to 455.7 (about 18%) with accuracy 28.7% in the efficiency table, a stated 2.9-point drop versus the full model in that table's numbers.

Surprising findings

• Chain-only RL-trained models often produce longer outputs, and Frugal Learning on Chain-Only data reduces accuracy more than on TopoAug data.
• The paper reports that Chain-only training is the best single topology overall (49% win rate), with tree and graph winning more than half of the remaining cases.

Limitations

• Relies on predefined topology types (chain, tree, graph) rather than learned topology induction.
• Evaluation sets are small (e.g., 100 to 500 samples per benchmark), and reported numbers are inconsistent across tables.
• The paper has not isolated the effect of topology diversity from general training effects.

Experimental setup

Datasets: MathVision (MATH-V) VLM_S2H OKVQA A-OKVQA LLaVA-150k-Instruct

Benchmarks: MATH-V VLM_S2H Geometry3K We-Math PolyMath SciBench LogicVista

Metrics: accuracy generated token length topology selection rate

Baselines: GPT-4o Qwen2VL-72B-Instruct Llama-3.2-11B-Vision-Instruct Phi-4-multimodal-5.6B-instruct LLaVA-v1.6-Mistral-7B MiniCPMv2.6-8B InternVL3-9B Chain-Only

Builds on: chain-of-thought prompting tree-of-thought graph-of-thought SimPO LoRA

Closest papers by meaning

• HopChain: Multi-Hop Data Synthesis for Generalizable Vision-Language Reasoning (NeurIPS 2026 · similarity 0.88)
• Unveiling Chain of Step Reasoning for Vision-Language Models with Fine-grained Rewards (NeurIPS 2025 · similarity 0.88)
• Zebra-CoT: A Dataset for Interleaved Vision-Language Reasoning (ICLR 2026 · similarity 0.88)
• ATLAS: Agentic or Latent Visual Reasoning? One Word is Enough for Both (NeurIPS 2026 · similarity 0.88)
• OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles (NeurIPS 2025 · similarity 0.87)
• Latent Chain-of-Thought for Visual Reasoning (NeurIPS 2025 · similarity 0.87)