It should be possible to edit model responses in chat interfaces, e.g. for editing the plan in Cursor and Claude Code.

RT @bjarnihaukur11: The funniest (unintentional) reward hack I saw while training my coding agent: it "rm -rf"'d the repo it was working on….

What do you want to know?.

Read more about GBPR, our experiments, challenges and future directions at

Our experiments on 1466 buggy programs show that GBPR can repair a large number of buggy programs to near-perfect accuracy. As the picture shows, gradient descent navigates the "correctness landscape", iteratively moving from buggy (high loss) to repaired (low loss) behavior.

How GBPR works:.1️⃣ Compile the symbolic program into a differentiable numerical form (e.g. a neural network). 2️⃣ Define a "correctness loss" based on the desired behavior. 3️⃣ Use gradient descent to adjust the numerical program's parameters, minimizing the loss. 📉➡️✅.

Our core idea with GBPR: reframe program repair as continuous optimization in a continuous program space. Instead of discrete token edits, we "steer" programs towards correctness by optimizing their numerical representations based on a correctness loss.

Program repair involves searching discrete symbolic spaces, lacking a direct way to optimize for program behavior. What if we could change that?. 📄 Introducing our new paper "Gradient-Based Program Repair: Fixing Bugs in Continuous Program Spaces"

RT @menhguin: I picked a great time to switch to Cursor for LaTex editing.

These and other updates available on

3️⃣Llama 4's underwhelming performance. Meta's release of Llama 4 was highly anticipated. However, its performance on RepairBench indicates that it does not outperform gpt-4o models and marginally improves over llama-3.1-405b, contradicting expectations based on LMArena scores.

2️⃣Gemini 2.5 Pro shows good progress from Google. Google's Gemini 2.5 Pro has demonstrated improvements over its predecessors, with a Plausible@1 score of 38.3% vs. 33.2% of the previous generation. Despite this, it still falls short of the Claude and DeepSeek models.

1️⃣Quasar-Alpha is a strong contender. Quasar-alpha, a stealth model available on OpenRouter, has made the news. With a Plausible@1 score of 40.5%, quasar-alpha is approaching the performance of leading models like Claude-3.5. The big question is: who is behind quasar-alpha?.

🔧 This week on RepairBench:. - Quasar-Alpha is a strong contender.- Gemini 2.5 Pro shows good progress from Google.- Llama 4's underwhelming performance

More details:

Just added claude 3.7 sonnet (non-thinking mode) to repairbench!. Strong improvement and almost on par with o3-mini and deepseek-r1. Wonder how far the thinking mode will land from the current top.

(probably good evidence that this new generation of models can handle more autonomy than the previous one).

claude 3.7 sonnet working so hard that even copilot thinks it's stuck

RT @mokita_j: 🔥sb-heists is now featured in this week’s @blockthreat newsletter!. 🔍91 reproducible exploits for 9 blockchain vulnerabilitie….

RT @martinmonperrus: OpenAI strikes back and reclaims first place 🥇 on the RepairBench leaderboard for automated bug fixing. https://t.co/….