Brief

last 24h

[2/2] 222 sources

Multi-source AI news clustered, deduplicated, and scored 0–100 across authority, cluster strength, headline signal, and time decay.

TOOL · LessWrong (AI tag) English(EN) · 3d

Looking for backdoors in Jane Street LLMs

A participant in Jane Street's LLM backdoor challenge shared their experience attempting to uncover hidden triggers in fine-tuned models. Initially, prompting strategies proved unsuccessful in revealing the backdoors. The challenge involved both a smaller, locally runnable Qwen2.5-7B-Instruct model and larger DeepSeek-V3 Mixture-of-Experts models accessed via API, with the latter proving particularly difficult to analyze. AI

IMPACT Details a novel approach to identifying vulnerabilities in large language models, potentially informing future AI security research.
RESEARCH · Hugging Face Daily Papers English(EN) · 5d · [3 sources]

CoSPlay: Cooperative Self-Play at Test-Time with Self-Generated Code and Unit Test

Researchers have developed CoSPlay, a novel framework for improving LLM code generation without relying on ground-truth unit tests. This training-free approach uses cooperative self-play to iteratively refine both generated code and its associated unit tests. By analyzing execution signals, CoSPlay prunes weak code and refreshes unreliable tests, leading to significant improvements in code generation accuracy and test quality. AI

IMPACT This framework offers a scalable inference strategy for competitive code generation, potentially reducing reliance on costly ground-truth data.