New studies show that fine-tuning on coding benchmarks does not generalize and that benchmarks often become training targets. Dedicated suites for clinical accuracy, long-novel summaries, and research tasks reveal calibration and hallucination gaps.
LongNovel is a multi-scale, bilingual Chinese-English benchmark for detecting hallucinations in long novel summaries. Built on 29 Chinese novels from 16k to 100k tokens and BookSum data, it identifies eight hallucination types. The test set was manua...
A Princeton-led study put Claude Opus 4.8 agents to work on unpublished NeurIPS 2026 research questions. The agents handled engineering tasks but lacked the judgment and creativity needed for open-ended research, and both papers were rejected. The fi...
A controlled clinical benchmark on gpt-4.1-nano shows 93.5% accuracy but imperfect calibration: confidence rises with evidence distance from the diagnostic boundary and falls with missing information, yet remains too high in moderate, conflicting err...
Fine-tuning on SWE-bench does not transfer capability to suites like Django or LiveCodeBench. The benchmark becomes a training target and stops measuring general ability. For self-hosted LLM operators, an inflated score distorts VRAM, quantization, a...
Optimizing an LLM on popular coding benchmarks does not improve general programming ability. A new Django-based suite shows rankings often fail to generalize, and fine-tuning on SWE-bench yields limited or no gains on other tasks. Evaluation must be ...
