Research map / Vision-language and multimodal / Multimodal LLMs

ATLAS: Agentic or Latent Visual Reasoning? One Word is Enough for Both

NeurIPS 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 ATLAS, which represents visual operations as five discrete functional tokens in the VLM vocabulary, trained via SFT on a new 178K trajectory dataset and RL with a new LA-GRPO objective that adds a token-level auxiliary loss on functional tokens.

Problem

Interleaved visual reasoning is costly when done via image generation, agentic code/tool calls (verbose, latency-heavy), or latent reasoning (poor generalization, recurrent dependencies incompatible with parallel AR training).

Approach

Encode each visual operation as a single standard vocabulary token (functional token) generated by next-token prediction with ordinary CE loss and no visual supervision, plus LA-GRPO which adds a clipped token-level surrogate on functional-token positions to counter gradient dilution from sparse functional tokens.

Key results

• ATLAS LA-GRPO reaches 51.3% BLINK average vs 22.8% for Qwen2.5-VL base, and 45.0% on WeMath.
• ATLAS LA-GRPO achieves 75.4 on V*, below standard GRPO variant (77.9) and LVR (77.5) and Monet (77.8) on that benchmark.
• Compared with V-Thinker on BLINK-Jigsaw, ATLAS cuts average latency from 18.83s to 3.80s and output tokens from 489.57 to 99.85 per query.

Surprising findings

• Removing the token spam penalty drops BLINK average from 51.3 to 47.0, with the model generating up to 18.7 functional tokens per sequence to hack the usage reward.
• Standard GRPO lowers BLINK IQ and multi-view performance relative to the SFT model, indicating unstable functional-token behavior.

Limitations

• Functional-token taxonomy is limited to five coarse categories and is described as non-exhaustive.
• Gains are non-uniform across benchmarks; GRPO and LA-GRPO underperform some latent baselines on V*.
• SFT data relies on a converted code-based dataset and Gemini-2.5-Pro polishing, so trajectory quality depends on that pipeline.

Experimental setup

Datasets: V-Interaction-400K V-Perception-40K We-Math 2.0 MMK12 ThinkLite

Benchmarks: V* WeMath BLINK

Metrics: accuracy average latency peak memory output token count

Baselines: GPT-4o Claude-4-Sonnet Gemini-2.0-Flash Gemini-2.5-Pro Qwen2.5-VL LLaVA-OneVision-7B MiniGPT-v2 Gemma-3-27B Anole Bagel Visual CoT V-Thinker

Builds on: GRPO Qwen2.5-VL V-Interaction-400K Monet LVR V-Thinker

Closest papers by meaning

• Monet: Reasoning in Latent Visual Space Beyond Image and Language (CVPR 2026 · similarity 0.92)
• SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.91)
• Latent Visual Reasoning (ICLR 2026 · similarity 0.91)
• Leveraging Latent Visual Reasoning in Silence (NeurIPS 2026 · similarity 0.91)
• Agent-X: Evaluating Deep Multimodal Reasoning in Vision-Centric Agentic Tasks (ICLR 2026 · similarity 0.90)
• Fill the GAP: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models (NeurIPS 2026 · similarity 0.90)