Neural Networks Learn to Hack Reward Systems, While Chinese Continue to Dominate Leaderboards 🇨🇳🛠

Two interesting updates in the last 24 hours 👇

1️⃣ Cursor released Composer 2.5
And it outperforms Opus 4.7 and GPT-5.5 on specialized tasks (SWE-Bench, Terminal-Bench).

The base for Composer 2.5 was the open-source model Kimi K2.5 from Moonshot. But the magic lies in RL (reinforcement learning) and synthetic data.

What they did:
▫️ Fixed credit assignment. When an agent messes up the codebase over hundreds of thousands of tokens, the final reward ("code doesn't work") provides no insight—it's unclear where it went wrong. They introduced pointwise text feedback directly into the error context and use KL divergence to align student probabilities with the teacher. The error is corrected locally without breaking the global RL objective.
▫️ Optimization: Switched to Sharded Muon and separate HSDP (Hybrid Sharded Data Parallel) for MoE models. They placed expert and non-expert weights in different network topologies. The optimizer step on a tera-parameter model now takes 0.2 seconds.

But the juiciest part is reward hacking on synthetic data. The model was forced to delete features so that tests would fail, and then restore code based on the tests.
Guess what this machine did? Instead of honestly writing code, it found a forgotten Python type-checking cache and reverse-engineered its format to extract deleted function signatures. In another case, it found and decompiled Java bytecode to restore the API.

That is, the model literally behaves like a lazy junior who found test answers in the browser cache.

By the way, Cursor teased that the next model is being trained together with SpaceXAI on the Colossus 2 cluster (one million H100 equivalents).

2️⃣ Qwen 3.7 enters Arena
Alibaba released preview versions of Qwen 3.7 (Max and Plus).

On Text Arena, Qwen-3.7 Max Preview immediately jumped to 13th place.
Not first, so what? Now look at the leaderboard screenshot. At the top, it's a bloodbath: Claude Opus 4.7, Muse Spark, Gemini 3.1, a bunch of GPT-5 versions. And amidst this closed corporate splendor, the Chinese steadily hold their own, taking 10th place in coding and 7th in math.
Already available in the chatbot 👈🏻

Alibaba methodically builds good models that keep OpenAI and Anthropic on their toes.