This work introduces a general steering technique called Semantic Overlays: small learned adapters applied at chosen prefill positions to a frozen model's residual stream that defends against the broad class of prompt injections that add instructions in untrusted context.
Abstract
Everything a language model sees is tokens. The serving stack knows what each span is -- user input, tool output, instructions -- but the model must keep track of that itself, and it can lose track or be confused: text can be written to read like anything. Prompt injection is a natural exploit of this phenomenon. By scrambling the model's understanding of span identity, an attacker can induce unwanted and potentially dangerous actions. Adding a non-textual channel to the model's input -- a way to communicate span identity beyond text -- mitigates this class of attack. We thus introduce a general steering technique called Semantic Overlays: small learned adapters applied at chosen prefill positions to a frozen model's residual stream. Laying an overlay over a span creates an out-of-band annotation channel that cannot be replicated by tokens. Unlike steering vectors, Semantic Overlays are trained, adaptable, and selectively applied. An overlay can encode complex semantics that reshape how the model perceives the marked span: asked to copy a code snippet under an overlay asserting that it is in a different programming language than it is, the model rewrites the snippet, faithfully, in the asserted language. Overlays are also composable, allow for transparent reading of underlying content, and can carry complex payloads -- including imperatives that the model will follow. An overlay which marks a span as"non-executable"defends against the broad class of prompt injections that add instructions in untrusted context. We report strong results on prompt injection benchmarks: SEP separation rises from 24.3% to 96.5% with utility unchanged (our scoring rule; we also correct a defect in the published grader), TensorTrust attack success rate falls from 34.8% to 6.6%, and all four PIArena attack families drop to 0% compliance, all while marked spans stay readable (92.5% exact copy rate).
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
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.
Autistic children experience disproportionately high rates of anxiety, yet common interventions such as cognitive behavioural therapy and traditional mindfulness practices may be less effective due to their cognitive and abstract demands. Yoga nidra, a form of guided meditation using concrete visualisations and breath awareness, may offer a more accessible alternative. This mixed methods study evaluated the feasibility and pilot efficacy of a co-designed six-week online yoga nidra intervention targeting anxiety in autistic children aged 8-14 years. Neurophysiological and psychological data were collected from 13 participants using parent- and self-reported measures of anxiety (ASC-ASD), intolerance of uncertainty (IUSC), and emotion dysregulation (EDI), alongside heart rate variability (HRV). Notably, the cohort included a minimally speaking child, which demonstrates the potential for the intervention to extend to autistic children with language support needs, a group who are often underrepresented in similar research. Results indicated no statistically significant changes in anxiety over time; however, medium effect sizes were observed in self- (Hedge's g = 0.55) and parent-reported anxiety (Hedge's g = 0.45), with three participants moving from clinically significant to non-significant anxiety levels post-intervention. Additionally, intolerance of uncertainty and emotion dysregulation demonstrated small to medium effect size reductions (IUSC, Hedge's g = 0.50; EDI-Reactivity, Hedge's g = 0.55; EDI-Dysphoria, Hedge's g = 0.25), suggesting potential benefits of yoga nidra in these areas. Greater attendance was significantly associated with self-reported reductions in anxiety (ASC-ASD-SR, r = .757, p < .05), but not the other outcome measures. Unexpectedly, HRV outcomes indicated reduced autonomic functioning post-intervention. Additionally, qualitative data from semi-structured interviews with three child participants and their mothers from the study were analysed using reflexive thematic analysis, revealing five key themes of their experience: Thinking About the Body is Hard; We Want More; Homework Sucks; Making Mindfulness Concrete; and On-Screen and At Home is Convenient. While the online format was appreciated for its accessibility, challenges with adherence and scheduling highlighted the need for more flexible delivery models. This study advances upon prior research by being the first to develop a co-designed yoga nidra intervention specifically for autistic children. Findings support the feasibility of yoga nidra as a complementary intervention for autistic children and suggest directions for future research, including larger trials and further co-design with the autistic community.
Tundi Loftus, Shu H Yau, Sophia Soares et al.· Research in Developmental Di...· 1 citation
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
FPGAgent is the first task-specification-to-executable HLS generation framework experimentally validated on a well-established benchmark, and the value of end-to-end validation is demonstrated.
Tianyu Wang, Wenjie Wang, Jianguo Yao et al.· 0 citations
Large Language Models show potential in their diagnostic accuracy and consequent ability to reduce clinician burden, and may provide the greatest benefit when used to optimise referral quality at source, improving both clinician and potentially LLM triage downstream.
K. Surendran, I. Aziz, Glyndwr Jenkins· Current Surgery Reports· 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.
