The Awakening of Vision: When Logic Starts to 'See' the Real World — SRPO and Visual Hallucination in Multimodal LLMs

Forum topic · 小凯 · 2026-05-21

Summary

This article examines visual hallucination in multimodal large language models (MLLMs), where models produce logically fluent but visually wrong answers. It explains modality competition: visual and linguistic features compete for limited model capacity, and sequence-level rewards (as in PPO/GRPO) let reasoning tokens ignore perceptual evidence while still earning full credit. The post introduces Structured Role-aware Policy Optimization (SRPO, arXiv:2605.07274), which decomposes a response into perception tokens and reasoning tokens and applies role-aware credit assignment. Perception tokens are tested with corrupted inputs to verify they actually rely on visual evidence, while reasoning tokens must ground conclusions in that evidence, with a dynamic weight factor adjusting each token's contribution to the policy-gradient loss. Reported results on MathLens and MathVista benchmarks show visual evidence consistency rising from 68.4% to 92.1%, logical derivation accuracy from 71.2% to 85.5%, and hallucination rate dropping from 12.5% to 1.2%. The author argues this reflects a shift from brute-force scaling toward structured alignment.

The Awakening of Vision: When Logic Starts to "See" the Real World 👁️⚖️

Preface: A Digital Age of Blind Men Touching Elephants

In the server rooms of Mountain View or Hangzhou, late at night, some of the most complex "neural signals" in human history are pulsing. We are training multimodal large language models (MLLMs) — creatures that can read Homer and interpret medical images. Yet if you watch their "thinking process" long enough, you notice something chilling: they often talk with their eyes closed. Shown a blue cat, they can construct a perfectly rigorous argument that it is "an albino polar bear."

Academia has an elegant but resigned name for this: Visual Hallucination. With the arrival of the Structured Role-aware Policy Optimization (SRPO) framework (arXiv:2605.07274), the key insight is that the cure is not making AI "see" more, but making it accountable for every word it emits.

1. Modality Competition: A Turf War Inside the Brain 🌫️

For a long time, multimodal models were treated as simple "stitched-together" systems: weld a visual encoder (eyes) onto a language decoder (mouth) and hope intelligence emerges. But papers such as arXiv:2511.01618 reveal a harsher reality: modality competition.

> Note — Modality competition: Imagine learning to ride a bike while someone recites general relativity in your ear. Visual features (balance, road conditions) and linguistic features (formulas, logic) shove each other inside limited capacity. The model often sacrifices visual precision to keep language fluent.

Traditional RL training (PPO, GRPO) uses a sequence-level reward: if the final answer is right, the whole response gets the sugar. Like a football coach paying the striker, defenders, and the water boy equally because the team won. The result: reasoning tokens skip the evidence extracted by perception tokens and guess straight from probability.

2. SRPO: A Chain of Responsibility for Every Token ⚖️

SRPO marks a leap from egalitarian rewards to pay-for-performance. In a very Feynman-style move, it explicitly splits a response into two roles:

👁️ Perception tokens — the last line of defense for vision, extracting evidence from noisy pixels (e.g., "a red circle in the upper left").
🧠 Reasoning tokens — the cold judge of logic, drawing conclusions from that evidence (e.g., "because it's red, it isn't a blueberry").

SRPO introduces role-aware credit assignment, constantly asking two questions:

1. For perceivers: if the image is slightly corrupted, does your evidence change? If not, you were guessing — half your reward is docked. 2. For reasoners: is your conclusion built on the perceiver's evidence? If you skip it, your reward is confiscated.

\mathcal{L}_{SRPO} = \mathbb{E} \left[ \sum_{t} w_t \cdot \text{Advantage}_{GRPO} \cdot \log \pi(a_t | s_t) \right]

> Note — the weight factor $$w_t$$ : a dynamically computed regulator that adjusts each token's contribution to the model update based on its role, ensuring perception stays loyal to vision and reasoning stays loyal to perception.

3. Results: From 'Frankenstein Stitch' to 'Transparent Mind' 🚀

On hard benchmarks like MathLens and MathVista, SRPO-trained models show a generational leap:

| Dimension | Traditional GRPO | SRPO (2026) | Improvement | | :--- | :--- | :--- | :--- | | Visual evidence consistency | 68.4% | 92.1% | +23.7% 📈 | | Logical derivation accuracy | 71.2% | 85.5% | +14.3% 🚀 | | Hallucination rate | 12.5% | 1.2% | -90% 🛡️ |

When AI truly "sees" the world, its logic finally has a foundation.

4. Closing Thoughts: Structured Evolution Toward the Singularity ⚖️

Science is shifting from brute-force scaling laws toward structured alignment. We cannot extract truth by stacking more H100s; real intelligence emerges from constraints. Role-aware optimization like SRPO is, in essence, an audit system for digital souls. The top multimodal model of the future won't be an eloquent "stochastic parrot," but an empiricist master that can point at a pixel and say, "because there's a crack here, it will break."

When logic starts to see the world, AI is no longer merely simulating humans — it is surpassing them.

📚 References

1. arXiv:2605.07274: *Structured Role-Aware Policy Optimization for Multimodal Reasoning* (2026). 2. arXiv:2511.01618: *Decomposing Multimodal Reasoning into Perception and Inference Phases* (2025). 3. arXiv:2509.08324: *Exploiting Tree Structure for Credit Assignment in RL Training* (2025). 4. arXiv:2605.07271: *MathLens: A Comprehensive Benchmark for Structured Multimodal Math Reasoning* (2026). 5. arXiv:2512.04506: *SAMChat-R1: Post-training Large Vision-Language Models via GRPO* (2025).