Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Official repo for "More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models" (ICLR 2026)
| Date | Stars |
|---|---|
| 2026-07-31 | 389 |
| 2026-08-06 | 389 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
<div align="center"> # **More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models (ICLR 2026)** </div> <p align="center"><i>A sober look at the pros and cons of multimodal reasoning with comprehensive findings, and a new RL method as a multimodal replacement of GRPO, achieving new state-of-the-art results.</i></p> <div align="center"> [](https://xytian1008.github.io/VAPO/) [](https://arxiv.org/abs/2509.25848) [](https://github.com/xytian1008/VAPO) </div> This is the official implementation of the paper 'More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models'. # News📰 * **`[2026/03/29]`:** 🔥 **We have released our training datasets.** * **`[2026/03/20]`:** 🔥 **We have released our code and models.** * **`[2026/01/25]`:** 🎉 **Our paper has been accepted to ICLR 2026!** * **`[2025/10/01]`:** 🔥 **We have released our paper [[Arxiv](https://arxiv.org/abs/2509.25848)].** # Key Findings🔍  🌟 **Longer reasoning does not guarantee better performance**: By breaking down the reasoning chains, we observe that the early stages of reasoning significantly enhance model accuracy. However, as reasoning continues, this performance gain gradually saturates and may even begin to reverse in later stages. 🌟 **The harder the model thinks, the worse the model sees**: Our error analysis reveals that prolonged reasoning is accompanied by an increase in perception errors, where the model incorrectly recognize or interpret visual details. This degradation in perceptual accuracy is a key factor underlying the negative effect of reasoning. 🌟 **The harms of reasoning are most evident in vision-heavy tasks**: In contrast to tasks with simple visual structures such as math, the adverse impact of reasoning on perception becomes more pronounced in vision-intensive problems involving high-resolution real-world images or perceptually elusive content.  🌟 **Encouraging models to look more often boosts reasoning**: We find that the drawbacks brought by reasoning stem from visual forgetting, where extended textual output leads the model to increasingly disregard visual cues. Encouraging models to attend to visual input effectively raises the upperbound of reasoning performance. # Methodology📖  We propose **Vision-Anchored Policy Optimization (VAPO)**, a simple yet effective policy gradient algorithm as a multimodal replacement of GRPO that explicitly steers the reasoning process toward visually grounded trajectories. The key idea of VAPO is to embed a sequence of visual anchors along the reasoning path. At each anchor point, the model's perceptual capability is probed by evaluating its responses to a set of primitive visual claims. Beyond standard outcome-based rewards such as accuracy and format, we introduce perception reward, which quantifies the model's overall perceptual grounding during reasoning by aggregating scores across all anchor points. # Main Results🗒️  ✅ **VAPO consistently improves accuracy across diverse benchmarks**: Our model outperforms recent reasoning models of the same scale on mathematical problems, achieving an average improvement of 2% (49.1% → 51.1%). The advantage is more pronounced on general-purpose tasks, where our method surpasses previous best results by 3.2% (59.9% → 63.1%), thereby establishing a new state of the art.  ✅ **VAPO fully releases the potentials of reasoning**: Compared with the baseline, our model demonstrates a more gentle decline in visual attention ratio, indicating that VAPO effectively
Excerpt of 9,160 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:9a2180eb2ef93fa3, desc:vision-language