Whether independently implemented training stacks can serve as differential oracles for a whole fine-tuning pipeline, rather than the operators and inference paths that prior differential testing targets, is studied.
Abstract
Neural network training has an oracle problem: a run can converge normally and yield a usable model while the software beneath it computes something other than specified. Almost all such work runs on one stack, so there is rarely anything independent to check against. We study whether independently implemented training stacks can serve as differential oracles for a whole fine-tuning pipeline, rather than the operators and inference paths that prior differential testing targets. We define a trajectory-level protocol -- a shared specification, cross-check points spanning arithmetic, model loading, data rendering and the learning trajectory, and a separation of independence of the stack, the orchestration and the language runtime -- and apply it to a LoRA adaptation of Qwen3-0.6B over 168,574 clinical question-answer pairs under PyTorch and under numbat, an independent framework written in Zig, driven natively and through its C interface from six languages. Across 42 paired evaluations spanning a full epoch the two stacks'held-out cross-entropy differs by 0.134% on average, and four implementations end the epoch within 0.15% of one another. The comparison exposed 17 faults that single-implementation development had missed, two of them notable for software engineering. The fault with the largest effect on the trained model lay outside the numerical kernels: a mismatch in how clinical text was rendered moved held-out loss 0.15, some 500 times more than the arithmetic faults found beside it. And four faults were reachable only from a language whose memory model differs from the first two implementations: a scheduler migrating work across threads, a collector blind to device memory, an ownership discipline needing a primitive the interface lacked. Implementation diversity has several axes, and the runtime is one.
Across medical benchmarks, MedLLM shows a pattern visible only at sub-billion scale: medical competence does not degrade uniformly under compression but splits by task type and dissociation is masked at 7B, where both capabilities are present, and surfaces only when capacity is scarce.
M. R. Rahman, Asim Ahmed, Mihan Mohagheghzadeh et al.· 0 citations
This tutorial presents OneEHR, an open-source toolkit that defines a unified experiment contract for modern EHR modeling and enables head-to-head comparison among conventional, neural, LLM-based, and agentic methods through a single configuration-driven interface.
Yinghao Zhu, Zixiang Wang, Lei Gu et al.· Proceedings of the 32nd ACM...· 0 citations
Agentic large language models are increasingly used across the genomic workflow, from variant calling to clinical interpretation, yet they are evaluated by accuracy alone, a single figure that cannot say whether a system is safe or where in the workflow a failure originates. We present ClawBench, a framework that attributes each outcome to the architectural layer that produced it across both halves of the canonical pipeline. Two design choices remove the confounds that make agentic genomics hard to evaluate: a temporally blinded truth set, in which every scored ClinVar label first became available only after the training cutoff of every model tested, and a fail-closed evidence contract that blocks evidence circular with the truth label. We score validity, safety, provenance and reproducibility, not accuracy alone, under a constraint gradient that relocates correctness from a model’s prior into executed, validated code. We show three things. First, dangerous misclassification is rare and model-invariant, a controlled precondition of the executed architecture rather than a frontier, while fabricated evidence is measurable and is neutralised by execution. Second, different variant classes are rate-limited by different layers: loss-of-function variants by the deterministic combiner threshold, and rare missense by evidence formation, where evidence acquisition is asymmetric and capped and strength assignment is a recoverable layer that naive strength-licensing prompts confound. Third, for variant calling the arms separate not on whether a model can plan a pipeline, which all do, but on trust properties, pinning, provenance, auditability and reproducibility, which climb monotonically toward validated execution; and a local open-weight model reproduces the safety result yet meets the structured-output and provenance contract far less often than frontier models, a conformance gap rather than a capability or safety gap. An end-to-end join attributes failures across the whole workflow, separating a missed call from a propagated genotype error from a correctly called but misinterpreted variant. ClawBench shows that apparently identical outcomes arise from distinct, independently measurable failure modes, and that trustworthiness in agentic genomics is a property of the pipeline architecture rather than of the model, providing a portable, contamination-resistant unit of attribution for the field.
This entry-level tutorial aims to equip healthcare professionals with the tools necessary to effectively integrate LLMs into clinical practice, ensuring that these powerful technologies are applied in a safe, reliable, and impactful manner.
Qiao Jin, Nicholas Wan, Robert Leaman et al.· Nature Protocols· 1 citation
Clinical language models can achieve strong in-hospital accuracy yet fail under deployment shifts because they exploit note-specific artifacts (e.g., templates, separators, boilerplate) that do not reflect patient state. We propose CAST (Concept-guided Artifact Suppression Tuning), an SAE-based framework for auditable clinical text classification. CAST uses Sparse Autoencoders to expose sparse, human-auditable features from intermediate Transformer activations, labels SAE latents with an LLM-assisted interpretation pipeline and ICD-10 retrieval constraints, suppresses verified artifact latents via residual subtraction during fine-tuning, and provides post-hoc per-concept attributions for auditing model decisions. On MIMIC-IV discharge-note mortality prediction, CAST improves over its corresponding fine-tuned encoder baselines and remains competitive with strong LLM baselines, while producing a feature-level audit trail of the clinical concepts that support each prediction and the artifact concepts suppressed during training.
This work proposes a measure of training data influence that does not require selecting a downstream task or validation set as the attribution target, and defines an example's influence by how much its gradient update reduces the squared distance to the final parameters of a given pretraining run.
Yuto Nishida, Hirokazu Kiyomaru, Yusuke Oda et al.· 0 citations
Related blog posts
MIT News · Artificial Intelligence· news.mit.eduAug 17, 2026
A USAF cadet and a Lincoln Laboratory researcher found AI chatbots can help nontechnical service members produce viable software applications for their unique problems.
