Auditing three rounds of rank-$32$ LoRA self-training on Qwen3-8B against a frozen control pushed through the identical pipeline, this work identifies seven measurement failures, each of which inverts a reported finding when its control is absent.
Abstract
Whether a language model has improved itself is increasingly judged not by mean accuracy but by which individual problems it gains and loses. Tracking these transitions means differencing two noisy estimates, leaving them vulnerable to measurement artifacts. Auditing three rounds of rank-$32$ LoRA self-training on Qwen3-8B against a frozen control pushed through the identical pipeline, we identify seven measurement failures, each of which inverts a reported finding when its control is absent. Several are standard practice. A ledger built on a single greedy decode manufactures capability changes on an untrained model, largely an artifact of inference batching; the expansion statistic separating acquisition from sharpening assigns that same model a rate of $0.280$. The natural threshold repair does not survive replication: estimated across the frozen comparisons such a design already contains, its null stays non-zero. We replace it with a per-problem exact test against a pooled baseline under false-discovery-rate control, which detects nothing on any held-out replicate and is unchanged under the multiple-testing rule, error rate and pool size. Applied to a ladder of arms matched in stream, volume and evaluation, the audit finds that external distillation improves problems the base model rarely reaches while three forms of self-training do not; a regression rejects this asymmetry as a by-product of distillation's larger overall gain ($p<10^{-8}$). On the far smaller set of problems the base model never reaches, the evidence is inconclusive, while self-training corrupts problems solved at baseline at rates well above the measured floor. Transition-level auditing therefore requires a separately measured null for every statistic it reports: nulls that cost no new experiments, built from baseline replicates a multi-arm study already owns, though not from as few as most possess.
LifeSciBench is introduced, a benchmark of 750 expert-authored tasks designed to evaluate whether language models can handle realistic life science research work, with each constituent task paired with a human expert-written rubric.
Amelia Liu, Andrew Ho, Anne Marie Droste et al.· bioRxiv· 2 citations
TestifAI, a deep learning testing framework for efficient and accurate estimation of robustness against combinations of perturbations, is proposed and partial model tomography is introduced, a novel approach to reconstructing model behaviour in a multi-perturbation space from tests that apply only a small number of perturbations.
Arooj Arif, T. Hartung, E. Botoeva et al.· 1 citation
A rapidly advancing precision-therapy pipeline-including antisense oligonucleotides to upregulate the intact allele, AAV-based gene replacement, CRISPR-mediated transcriptional activation, epigenetic modulators, and rational pathway-targeted small molecules-offers realistic prospects for disease modification.
Two language-model-based strategies are proposed for semantic code document segmentation, including a line-by-line approach that classifies each line of code separately before grouping the results into functional units, and a range-based approach that aims to directly determine groups of code lines from the input.
Abdelhalim Hafedh Dahou, A. Scherp, Sebastian Kurten et al.· Proceedings of the 2026 ACM...· 0 citations
The results suggest that bitsandbytes 4-bit quantization can impose an additional cost on applications relying on long, updatable, semantically dense contexts, even when aggregate benchmark accuracy appears largely unaffected.
Background Ambient AI documentation tools, known as scribes, are entering routine clinical practice at scale, but the evidence comparing the notes they produce against clinician-written notes is dominated by single-site, single-language studies that rely on human review to find errors, a method known to miss most documentation errors. Methods We conducted a paired simulation across five countries and languages (Cambridge/English, Barcelona/Spanish, Milan/Italian, Paris/French, Cologne/German; 385 paired consultations, 770 notes). From each actor-performed consultation, an AI scribe (Heidi) and a junior-to-middle-grade clinician independently produced a note. Notes were scored on the PDQI-9 by evaluators blinded to authorship. Documentation errors were identified by two methods of deliberately different sensitivity - clinician adjudication, and a calibrated automated reviewer externally validated against a blinded ten-clinician panel - then graded for clinical risk by a three-model panel. The co-primary outcomes were PDQI-9 total and Critical+High error burden, the latter reported under both detection arms. The analysis plan was registered before any pooling across sites. Results AI notes scored higher than clinician notes on the PDQI-9 (40.6 vs 35.6; difference +5.08, 95% CI 4.6-5.6; Cohen dz=0.55), consistently across all five sites (dz 0.41-0.75), and were less dispersed (5.7% of AI vs 27.8% of clinician notes fell below the study pre-specified low-score threshold (<32)). On the principal safety outcome - the paired probability that a note carried [≥]Critical+High error - clinician notes were affected more often under both detection arms: 61.0% versus 24.4% by the calibrated reviewer (relative risk 2.50, 95% CI 2.09-3.00) and 21.8% versus 6.2% by clinician adjudication (relative risk 3.50, 95% CI 2.32-5.27). The difference was largest for omissions. Unaided clinician review identified roughly 12% of the errors the calibrated reviewer retained, and a smaller fraction in AI notes than in clinician notes. Conclusions In this simulation, AI-generated notes scored higher on documentation quality, varied less, and carried fewer clinically significant errors than notes written on the same consultations by junior-to-middle-grade clinicians. The magnitude of the safety difference depends on the sensitivity of error detection, so we report both detection regimes and bound rather than point-estimate the absolute error rate. Extension to live practice, consultant-authored documentation, and notes as filed after clinician editing remains to be established.
H. Bergman, V. Liu, B. Austin et al.· medRxiv· 0 citations
A new method for surgically removing training examples from a model reveals that as datasets grow, the link between what a model learns and what it produces dissolves.
