AI Box

The official Xaccount of Al Box group, Gaoling School of Artificial Intelligence, @RenminUniv

[AI Box @RUC_AIBox]

📄 Paper: https://t.co/rtSbr7A107 💻 Code: https://t.co/Lx9bbyeJ1C
#CVPR2026

[AI Box @RUC_AIBox]

Our CVPR 2026 paper is out! 🎉

VLMs reason in multi-step chains, but RL (GRPO) only rewards the final answer, so the model never learns which step actually hallucinated.

Perceval fixes this with token-level, perception-grounded rewards. 🧵👇

[AI Box @RUC_AIBox]

Results on Qwen2.5-VL (3B & 7B) vs GRPO: +~4% on visual search and +~3% on math & chart. Perception-only training generalizes to math/chart reasoning, and it rivals tool-using models like DeepEyes with no external tools.
• +~4% visual search, +~3% math & chart
•Results :

[AI Box @RUC_AIBox]

From one document, rebuild the whole repository.

That is the threshold long-horizon code agents must cross.

DeNovoSWE is a step toward agents that can engineer software systems, not just edit code.

[AI Box @RUC_AIBox]

Code agents are entering a new phase.

Beyond single-bug fixing.
Beyond patch-level SWE data.
Meet DeNovoSWE: an open-source long-horizon dataset for generating whole repositories from scratch.

[AI Box @RUC_AIBox]

The next stage of code agents is not just faster issue fixing.

It is document understanding, architecture planning, module organization, API implementation, and full-repository construction.

DeNovoSWE turns this goal into trainable, verifiable, scalable data.

[AI Box @RUC_AIBox]

Frontier Harnesses Need Guidance Too

The Codex comparison makes this sharper.

With the same GPT-5.4 (xhigh) model and harness, a SearchSWE-style prompt raises AVG from 48.48 to 56.65.

[AI Box @RUC_AIBox]

Can current code agents survive beyond single-repo bug fixing?

BeyondSWE: 500 real-world tasks from 246 GitHub repos, covering cross-repo issues, domain-specific fixes, dependency migration, and doc-to-repo generation.

📑 https://t.co/nq5akf1m32

[AI Box @RUC_AIBox]

Search Helps, But Unevenly

Search helps: 7/8 models improve, and 20/32 task-model pairs improve.

But 31.2% still regress. External information access is useful, not automatically reliable.