The paper reports a structured case study in which six frontier language models were asked to match handwritten exam answers to the ages of the writers. Three of the models performed at chance on the matching task while still reporting high confidence, and the youngest writer’s answer was misattributed to an adult in up to 90% of trials. Removing the multiple‑choice scaffold and changing institutional framing further degraded accuracy, while a single audience‑awareness instruction improved performance for one model but had no effect on the weakest models.
arXiv:2607. 28639v1 Announce Type: cross Abstract: We show that knowledge distillation in small instruction-tuned language models has asymmetric effects on bias.
A 0.6B language model consistently answers YES to 1,200 logical tests, yet its behavior shows no discrimination. Linear probes reveal the correct verdict with high AUC (0.96) and transfer to unseen structures, but a single scalar readout fails due to a saturated decision threshold offset by +4.6 σ. Adjusting this threshold restores behavior accuracy from 50 % to 81 % and improves higher‑scale models, demonstrating that miscalibrated readouts, not hidden knowledge loss, drive performance gaps.
arXiv:2609.13773v1 Announce Type: new Abstract: LLM-as-a-Judge evaluators are increasingly used to score open-ended generation, yet a judge's correlation with human ratings on its development set may...
arXiv:2603.11413v4 Announce Type: replace-cross Abstract: A recent Nature Medicine study reported that ChatGPT Health under-triages 51.6% of emergencies and concluded that consumer-facing AI triage p...
arXiv:2608. 11415v1 Announce Type: cross Abstract: Large language models are being proposed as agents in scientific workflows, in domains where no downstream verifier exists.
The study evaluates zero‑shot essay scoring using sub‑3B open language models that run locally on a single 8 GB consumer GPU. Four instruction‑tuned models (Qwen2.5‑0.5B, 1.5B, 3B and SmolLM2‑1.7B) were tested on all eight ASAP‑AES prompts, comparing rubric‑decomposed versus holistic prompting, different aggregation methods, and trait‑mapping strategies. Results show rubric‑decomposed prompting consistently outperforms holistic prompting, trait‑mapping is sensitive to calibration, and longer essays reduce error, yet the best local configuration (macro QWK 0.388) still falls short of human agreement and a length‑only baseline, suggesting these models are best suited for formative, human‑supervised feedback.
The study evaluates how clinical language models perform when patient narratives are presented in different voices—clinical, socioeconomic, and cultural—using 1,000 MedQA cases paired with three retellings each. All 12 tested models show a significant drop in accuracy (about 26% more errors) when the patient speaks in a marked voice, a penalty that persists even when the original vignette is included verbatim and is not explained by measured retelling properties. The authors find that rewriting the patient’s account into a clinical vignette before answering can halve this accuracy gap, though structured schema pipelines and additional reasoning tools offer limited improvements.
The paper investigates whether language models’ self-reported confidence is meaningful without additional training. By evaluating three training‑free signals—direct verbalization, post‑hoc probability estimates, and agreement across multiple generations—on 100 TriviaQA questions, the authors find that direct verbalization alone achieves high AUROC scores (0.956 and 0.937) for correctness prediction, while agreement-based methods perform noticeably worse. Re‑eliciting confidence for the same answers shows modest score shifts and occasional decision flips, and an audit of biography claims reveals only a small confidence gap between supported and contradicted statements.
The paper "Fluent but Fallible: Benchmarking Hallucination and Factual Reliability in Large Language Models" evaluates ten advanced LLMs on a benchmark of 70 academic prompts across four complexity levels. Using five metrics—Accuracy, Precision, Recall, Response Deviation Index, and Hallucination Index—reviewers found that Perplexity and DeepSeek V3 achieved the highest overall accuracy, Gemini balanced recall and performance, LLaMA 3 resisted hallucinations, and Falcon had perfect precision but low recall and high deviation. The study highlights a trade‑off between response completeness and factual reliability, especially as prompt complexity increases.
arXiv:2606. 16890v1 Announce Type: cross Abstract: Aggregate accuracy benchmarks conceal a systematic structure in how large language models fail at electronic health record (EHR) question answering: questions requiring more inferential steps produce disproportionately more errors.
arXiv:2607. 20526v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly deployed in settings where fluent but incorrect answers can be costly.
The study investigates whether large language models (LLMs) that allocate extra computation during inference—termed reasoning models—reduce classic decision biases compared to their non‑reasoning counterparts. Using 30 vignettes covering six cognitive biases and varying token budgets up to 8,192 tokens, the authors find that reasoning models are not less biased, and increased deliberation does not reliably diminish bias magnitude. Only anchoring showed a bias in the human direction, while other biases either remained unchanged or moved further from human patterns, suggesting that test‑time reasoning does not guarantee rationality.