Latest posts tagged with #CodeLLM on Bluesky
Activation Steering Improves CodeLLM Type Prediction Across Languages
Activation steering revives hidden reasoning in LLMs, restoring accurate type predictions for both Python and TypeScript after syntactic edits. The paper was revised Oct 3 2025. Read more: getnews.me/activation-steering-impr... #codellm #typescript
Code Language Models Can Learn to Ask Clarifying Questions
ClarifyCoder, a new code‑LLM framework, asks clarification questions on ambiguous prompts, achieving a 63% communication rate and a 52% good‑question rate. Read more: getnews.me/code-language-models-can... #clarifycoder #codellm #ai
V-GameGym Benchmark Expands Visual Game Generation Testing for Code LLMs
The new V-GameGym benchmark, released on 24 Sep 2025, includes 2,219 game samples across 100 thematic clusters to evaluate code LLMs on visual and interactive quality. Read more: getnews.me/v-gamegym-benchmark-expa... #vgamegym #codellm #visualai
FreqRank: Detecting and Localizing Malicious Code LLM Outputs
FreqRank spots hidden backdoor code in LLMs, ranking malicious output in the top five suggestions for 98% of cases and achieving 86.6% attack success across tasks. Read more: getnews.me/freqrank-detecting-and-l... #freqrank #codellm
Study Shows Code Semantics Offer Limited Gains for LLM Reasoning
An arXiv study reports that adding execution‑trace data to Code LLM prompts gives only marginal gains in supervised fine‑tuning and no significant improvement at test time. Read more: getnews.me/study-shows-code-semanti... #codellm #executiontraces
OpenCoder releases not just weights but the entire recipe—dataset, pipeline, ablations. That’s openness you can audit. #OpenSource #CodeLLM #AI
