The author reports significant regressions in tool-calling behavior with the latest Anthropic models, likely due to aggressive reinforcement learning on a closed-source harness. This issue causes failures when tool declarations are slightly off, a problem that did not exist in older model versions.
Background
Recent advancements in Large Language Models have focused heavily on improving function calling capabilities, often through reinforcement learning from human feedback (RLHF) or similar techniques. However, optimizing for specific internal environments can sometimes lead to unexpected fragilities in general-purpose API usage.
- Source
- Lobsters
- Published
- Jul 5, 2026 at 05:51 AM
- Score
- 6.0 / 10