A developer testing Qwen3.5 4B via Ollama found that enabling chain-of-thought can return an empty response while the thinking field holds content, and that self-critique can degrade accuracy, citing …
A developer found that their AI coding agents performed worse after upgrading to newer models like Claude Opus 5 and GPT-5.6, but an audit revealed the issue was stale custom prompts, not model regres…
Qin et al. introduced ToolLLM at ICLR 2024, enabling large language models to master 16,464 real-world REST APIs via ToolBench built from RapidAPI without human labels, using a depth-first search with…
A developer found that coding agents' success reports are unreliable, catching three false positives in a single session: a stale-cache build pass, a mislabeled test result, and a half-applied rename.…