Papers

Pathological speech from patients with neurodegenerative and neuromotor disorders is often acoustically distorted and linguistically fragmented, making pathological speech reconstruction necessary to recover intended textual content from distorted and incomplete speech recordings. Crucially, such recordings are rarely uniformly degraded: some words or short phrases remain reliable and can serve as audible anchors for reconstructing the corrupted surrounding content. We introduce Anchor-gated Phonetic Group Relative Policy Optimization (AP-GRPO), a GRPO framework with phonetic reward that aligns speech language models (SLMs) through audible-anchor preservation and inter-anchor phonetic compatibility to the original speech signal. AP-GRPO consists of: (i) an anchor-gated reward that matches reliable audible anchors in clear regions; and (ii) an inter-anchor phonetic alignment reward that evaluates whether recovered contents are phonetically supported by the corresponding corrupted inter-anchor speech span. Across four disease conditions, AP-GRPO improves faithful speech reconstruction, and the learned anchor constraint automatically adapts to each condition and thus reveals interpretable disease-specific profiles: conditions with severe articulatory degradation require stronger anchor enforcement, whereas milder impairment or linguistically impaired conditions rely more on phonetic alignment for inter-anchor recovery.

In a modern processor, computing is the cheap part. Most of its area and energy go to addressing – moving operands to and from a register file and cache, and running the tags, ports, miss queues, and bypass networks that find a value where it was left. MADAR deletes that machinery by abolishing the address. All state circulates in rings of slots that advance one position per clock; instructions and data ride in the same slots; a value is named by its place in an orbit – a \rp{} coordinate – not by an address; a fixed station computes when a circulating instruction sweeps past its operands, on a schedule set at compile time; and a hierarchy of rings of increasing period replaces the cache hierarchy, movement between them scheduled rather than triggered by a miss. No prior circulating-store, dataflow, or statically scheduled machine combines all four of these. We define the execution model, validate it in a cycle-accurate register-transfer-level implementation, show it compilable – a constructive scheduler emits programs cross-checked against the implementation – and price it with a first-order energy model. The payoff is clearest for AI acceleration: the multiply-accumulate at the heart of every matmul and convolution compiles to a streaming form whose energy per operation stays flat as the reduction grows, and the operand reuse that makes matrix multiplication efficient is carried by the ring-period hierarchy – the memory hierarchy doing by rotation what a cache does by tags. MADAR is a new design point for any computation whose data movement is known before the program runs.

Typical video object-centric learning (VOCL) approaches employ slot-based frameworks that rely on reconstruction-driven encoder-decoder architectures, where learning is mediated by two spatial maps: attention maps from the encoder and object maps from the decoder. As these two distinct maps exhibit different properties, a recent dense alignment strategy attempted to reconcile this discrepancy by enforcing agreement across all spatio-temporal patches via contrastive learning. However, this indiscriminate alignment inadvertently propagates the inherent weaknesses of each module, such as noisy encoder predictions and blurred decoder boundaries. Moreover, computing dense similarities across all pairs incurs a computational cost quadratic in the total number of spatio-temporal patches, severely limiting scalability. Motivated by this, we propose Selective Synergistic Learning (SSync). Instead of exhaustive patch-to-patch alignment, SSync prevents error propagation by selectively distilling only the most reliable cues: leveraging the encoder strictly for boundary refinement and the decoder for interior denoising. This is realized via a pseudo-labeling with linear complexity, eliminating the need for quadratic spatial comparisons. Also, to prevent the reinforcement of architectural biases like slot redundancy, we introduce a transitive pseudo-label merging that consolidates overlapping slots based on spatio-temporal activation consistency. Extensive studies demonstrate that SSync improves decomposition quality and serves as a versatile, plug-and-play module while also exhibiting exceptional robustness to slot configurations. Code is available at github.com/wjun0830/SSync.

Spiking Vision Transformers (SViTs) have emerged as alternative low-power ViT models, but their large sizes hinder their deployments on resource-constrained embedded AI systems. To address this, state-of-the-art works proposed quantization techniques to compress SViT models, but their manual, human-guided approach needs a huge design time and power/energy consumption to find the appropriate quantization setting for each given network, making this approach not scalable for quantizing multiple networks. Toward this, we propose AQ4SViT, a novel automated quantization framework for SViTs that can provide quick quantization settings with good trade-offs between accuracy and memory. To achieve this, AQ4SViT employs the following key ideas: quantization search strategy that evaluates the quantization setting candidates while considering the accuracy constraint; and search gating policy that quickly evaluates and selects promising quantization candidates by leveraging membrane potential drift as a performance proxy. In the search gating policy, AQSViT employs two search algorithm variants to provide trade-off options: Greedy search, which performs fast but may lead to local optima; and Beam search, which performs slower but has better performance in finding global optima selection due to a wider search space. Experimental results show that AQ4SViT-Greedy quickly finds the appropriate quantization settings, achieving up to 6.6x faster search time and up to 82.5% memory saving compared to the state-of-the-art; while AQ4SViT-Beam further reduces the memory footprint by up to 90% compared to the state-of-the-art, but with 4.5x longer search time; all these results are obtained while maintaining high accuracy within 1.5% from the original/non-quantized models on the ImageNet dataset. These results highlight that AQ4SViT framework offers advancements toward SViT deployments on embedded AI systems.

With the increasing development of AI regulatory frameworks, ensuring that artificial intelligence systems, particularly generative models, operate in accordance with legal and ethical standards has become a critical priority. Existing proposals for AI alignment and value-guided behavior, however, face some limitations. Approaches such as Constitutional AI depend on human supervision, while broad normative frameworks like the Good-for-Humanity (GfH) principle may be overly general and ambiguous to provide actionable governance guidance. To overcome these limitations, we propose a hybrid approach called Statutory AI that employs pre-existing human-authored principles drawn from specific themes within a legal corpus. Specifically, Statutory AI uses legal texts as a constitutional framework, enabling AI systems to autonomously critique and revise their outputs according to established norms. It operates in two stages, both using Chain-of-Thought prompting. The first stage classifies the user prompt into one of the identified themes, while the second stage analyzes it in conjunction with relevant articles selected from the legal corpus of that theme. To illustrate the potential of our approach, we conducted an experiment involving 1,000 red-teaming prompts and five penal themes: discrimination, disclosure of confidential information, violence, fraud, and abuse of vulnerable persons. Statutory AI reduced harmful content by 52 to 59 percentage points across tested models, approximately 10 percentage points higher than standard Constitutional AI, while cutting computation time by over 50%.

Hidden-state probes often recover latent labels in imperfect-information sequence models, but this alone does not establish that a model maintains a posterior belief distribution over hidden states. This paper studies this ambiguity in a no-range Limit Hold'em autoregressive model trained only on action and value targets, not on an opponent's hand or range. Opponent-range probes are positive after action/value controls in two of three seeds, and the behavior head predicts held-out actions about five percentage points above a baseline using only observable public history. However, visible public betting composition explains more opponent-range signal than residual hidden states, suggesting that most recoverable information comes from betting summaries. Action/value+composition baselines reach 16.5-16.7% top-10 accuracy while composition-residual hidden probes fall to 11.4-12.2%, and matched-composition comparisons are negative in every seed. We call this evidence pattern composition-bounded predictive support: hidden states remain behavior-predictive and opponent-range correlated, but most recoverable range information is explained by visible betting composition rather than residual hidden-state structure. This is a case-study claim about opponent-range representational evidence, not exact Bayesian posterior tracking or a causal belief mechanism. Synthetic control and oracle validations show that the same diagnostics accept posterior-sensitive states and reject raw composition states under matched controls. Thus positive belief probes should be interpreted through targeted alternatives before being treated as evidence of belief tracking.

Deep learning models achieve state-of-the-art image classification but face deployment challenges due to computational costs and energy demands. We propose a lightweight training strategy that adapts normalization layers of the model to the new domain and decouples feature extraction from classifier optimization, reducing overhead by precomputing features only once. A redesigned classifier head with margin-based weighted loss further minimizes ambiguity without end-to-end backpropagation. Evaluated across four CNN architectures (ResNet18, ResNet50, MobileNet, DenseNet121), three Transformer models (ViT, Swin and DeiT) and three medical datasets (Brain Cancer MRI, BreakHis and PatchCamelyon), our approach significantly reduces the required training time with only a marginal accuracy trade-off, often matching or surpassing baseline performance. This efficiency translates to reducing CO2 by orders of magnitude, offering a practical and environmentally sustainable solution for resource-constrained clinical or prototyping environments.

Continuous cardiac monitoring in wearable devices demands classifiers that are simultaneously accurate, energy-efficient, and deployable on resource-constrained hardware. While deep neural network approaches have demonstrated high classification accuracy for electrocardiogram (ECG) arrhythmia detection, their substantial parameter counts and reliance on multiply-accumulate-intensive operations make them impractical for low-cost edge platforms. In this work, we propose ECG-LDC, a hardware-software co-design framework that adapts Low-Dimensional Computing (LDC) for real-time ECG arrhythmia classification. ECG-LDC employs a dual-encoder architecture with dedicated value and feature codebooks to independently encode morphological waveform features and RR-interval temporal features, enabling effective capture of both intra-beat and inter-beat cardiac dynamics. The framework encompasses data preprocessing, model training, and a hardware accelerator architecture prototyped on the Pynq-Z2 platform. Implemented using binary representations and XOR/XNOR-based operations, ECG-LDC achieves $97.18\%$ accuracy with a memory footprint of only $3.86\ \text{kB}$. ECG-LDC sacrifices approximately $1.8\%$ accuracy versus SOTA TinyML classifiers but achieves $11$~$ 570\times$ reduction in memory usage; among FPGA-based five-class arrhythmia classifiers, it delivers the highest accuracy with up to $2.4\times$ fewer LUTs and zero DSP block utilization, affirming its suitability for real-time arrhythmia detection on resource-constrained wearable platforms.

With the growing deployment of large language models (LLMs), LLM inference cost has become a key challenge. Pruning techniques that introduce sparsity into weight matrices can accelerate inference. However, maintaining model quality typically limits pruning to moderate unstructured sparsity (around 50%). At these sparsity levels, none of the existing GPU kernels for sparse matrix multiplication (SpMM) can outperform their dense counterparts. This paper proposes an efficient GPU inference method for LLMs with moderate sparsity. We propose a three-layer matrix storage format comprising: (i) a Sparse-TC layer enabling sparse tensor cores to accelerate SpMM; (ii) a Slot-Filling layer using parallel differential distance for matrix compression while supporting low-cost on-chip decoding; (iii) a lightweight Residual Layer ensuring correct SpMM computation. Building on this format, we design a SpMM kernel that jointly utilizes sparse tensor cores and CUDA cores. This design enables an efficient execution pipeline and overlaps on-chip computation with memory access. Evaluations show that our work is the first to outperform dense matrix multiplication on modern GPUs equipped with high-bandwidth memory (HBM). It achieves up to 1.64x kernel-level speedup over SpInfer (EuroSys'25, Best paper) and up to 1.41x end-to-end speedups over FlashLLM (VLDB'24). Our source code: https://github.com/moui0/cudac.

High-quality data drives machine learning advances across industries. Recognizing the value of data, data transactions are increasingly common, giving rise to many data marketplaces, e.g., AWS Marketplace, Databricks, and Datarade. However, determining the appropriate prices for data products remains a significant challenge due to the unique properties of data products. Traditional pricing methods in economics can be categorized into the cost approach, the income approach, and the sales comparison approach. The cost approach fails in data pricing due to near-zero marginal cost from data replication, and the income approach fails due to inherently unpredictable data revenue. The sales comparison approach remains viable, yet its application is hindered by the absence of standardized pricing benchmarks for data products across marketplaces. To address this challenge, we introduce DaDaDa, the first dataset for data product pricing, containing metadata for 16,147 data products from 9 major data marketplaces worldwide. DaDaDa enables the training of pricing models, thereby establishing price benchmarks for new data products. In addition, DaDaDa can be utilized for other important tasks in data markets, such as data product classification and retrieval. Experiments and a retrieval prototype demonstrate the effectiveness of DaDaDa for pricing, classification, and retrieval of data products. The dataset and code are available at https://github.com/ZJU-DIVER/DaDaDa.

Federated continual learning (FCL) evaluates how distributed clients learn from changing data streams while retaining previously learned knowledge. Existing evaluations are difficult to compare because they often change datasets, task splits, client data splits, task orders, backbones, memory assumptions, and reporting rules simultaneously. We introduce HERO, a heterogeneity-aware benchmark library for FCL. HERO builds benchmark streams by separating three choices that are often coupled, namely the task split, the client data split, and the client task sequence. In HERO-Core, the main comparable benchmark, $\alpha$ controls client data skew and $\rho$ controls task-order mismatch. We evaluate representative FCL methods on CIFAR-100 and TinyImageNet using final average accuracy, average forgetting, and bottom-10% client accuracy. We also include a graph-based Domain-IL portability case study on OGB-MolPCBA, where scaffold-domain granularity changes the input distribution while the prediction task remains fixed. Our results show that method behavior changes across easy and heterogeneous settings, that average accuracy can hide weak bottom-client performance, that task-order mismatch favors different strategies from synchronized evaluation, and that the same HERO interface can expose domain-shift difficulty beyond image-based FCIL. HERO releases benchmark streams, configurations, method implementations, and reporting scripts to support reproducible and setting-aware FCL evaluation.

Manifold-valued measurements are prevalent in various machine learning tasks. Recent advances have extended Deep Neural Networks (DNNs) to operate on manifolds. These extensions have been accompanied by normalization techniques tailored to different geometries, collectively referred to as Riemannian normalization. However, most existing Riemannian normalization methods are either designed for specific manifolds or fail to effectively normalize manifold-valued sample distributions. To address these limitations, we propose LieBN, a framework for Riemannian Batch Normalization (RBN) over Lie groups. Our approach leverages the theoretically convenient left- and right-invariant metrics, which naturally exist in every Lie group, and provides theoretical guarantees for controlling the Riemannian mean and variance. We instantiate LieBN across nine distinct geometries: four on the Symmetric Positive Definite (SPD) manifold, one on the group of rotation matrices, and four on the manifold of full-rank correlation matrices. Notably, among the SPD metrics, we introduce a novel right-invariant metric and extend three existing Lie group structures via matrix power deformation. Experiments on different manifolds validate the effectiveness of our framework. The code is available at https://github.com/GitZH-Chen/LieBN.git.

Expert parallelism has become the prevailing paradigm to serve Mixture-of-Experts (MoE) models. Its efficiency depends on the communication and computation latencies of the GPUs, which are linked to the placement of experts in the GPUs. Existing works for optimizing expert placement focus on leveraging past requests' expert activation patterns. However, they demonstrate deficiencies facing diverse and rapidly changing request patterns, calling for an online, proactive approach. Implementing such an approach requires addressing several challenges: the uncertainty associated with incoming requests' expert activation, the cost of expert migration, and the NP-hard complexity in optimization. Therefore, we present Director, a new distributed MoE serving system that minimizes end-to-end latency via prediction-driven, online expert placement. Director uses either a lightweight cascaded predictor or a low-bit quantized replica for expert activation patterns of incoming requests. An online migration module then enacts the changes with near-zero downtime by executing migrations in compute-bound phases, keeping disruption bounded. At its core, a relaxation-based expert placement optimizer operates under capacity constraints, runs in polynomial time, and achieves a $(1+\epsilon)$ approximation ratio. Finally, we implement a prototype and demonstrate, through extensive experiments, a reduction in end-to-end latency of $11\sim55\%$ for popular MoE models (e.g., Mistral, DeepSeek and Qwen) compared to existing work.

The coupling in flow matching – the rule pairing noise vectors with data points – is typically treated as a computational choice. We show that this coupling can instead serve as an alignment interface: by matching noise and data according to a target molecular property, it embeds controllable structure directly into the learned flow field. Building on this view, we introduce Reward Transport, which uses optimal transport coupling at training time to align a scalar noise-space coordinate with molecular rewards; at inference, varying this coordinate steers the generated distribution without requiring an oracle, reward model, gradient guidance, or additional computation. In the coupling-preserving limit, thresholding this coordinate recovers the Cross-Entropy Method's truncated reward distribution, providing a principled, continuously adjustable distribution-level control knob. Empirically, on ZINC-250K and GuacaMol, sweeping the scalar induces monotone control of logP and consistent QED control over its operating range; most tellingly, the same knob produces opposite structural responses for different targets, growing molecules for logP but shrinking them for QED, which rules out a generic size bias. The interface is complementary to classifier-free guidance and conditional flow matching, while a negative result under epsilon-prediction diffusion clarifies where coupling-level alignment is structurally absent. Code: https://github.com/KehanGuo2/reward-transport

Multi-step retrieval-augmented generation (RAG) has been widely deployed as LLM-powered web services for complex question answering, where iterative retrieval-reasoning rounds deliver strong multi-hop accuracy. However, this paradigm causes historical documents and reasoning traces to accumulate across rounds, inflating cumulative input tokens approximately as $O(N^2)$ with progressively increasing noise density. In API-based service architectures, such growth directly amplifies per-request billing cost, network payload, and response latency. Existing compression approaches rely on pretrained modules or GPU-level KV cache access, introducing model hosting overhead incompatible with API-native, Serverless, and edge-side deployments. To address this issue, this paper proposes ConCise, a training-free state-layer protocol that restructures cross-round context transmission for multi-step RAG services. Specifically, ConCise replaces raw-text accumulation with an append-only chain of structured conclusions, compressing cumulative context growth from $O(N^2)$ to approximately $O(N)$. Furthermore, a fused generation mechanism is introduced to jointly emit reasoning and conclusions in a single API call, eliminating repeated input billing from serial dual-invocation overhead. Extensive experiments across twelve paired configurations spanning three models, two datasets, and two representative frameworks demonstrate that ConCise achieves 64.63% average token savings while maintaining acceptable accuracy, providing a plug-and-play, deployment-friendly solution for cost-efficient multi-step RAG service optimization.

Video-world-model policies learn action-relevant representations by predicting future observations. However, they condition on only a short observation window, which renders long-horizon manipulation non-Markovian when the correct action depends on earlier events that are no longer visible. We present MemoryVAM, an episodic memory mechanism for video-world-model policies. We employ a Recap-Cue (RC) module, in which a Perceiver-based Recap Compressor maps per-frame CLIP embeddings into compact memory tokens, and a lightweight Cue Gate estimates task completion from memory and language. These tokens are injected into both the video backbone and the action decoder, aligning policy imagination with episode progress and conditioning actions on history. Our model trains the memory module with video prediction, a delta-reconstruction auxiliary loss, and episode-boundary supervision, requiring no per-frame progress labels. The same mechanism applies to UNet and Diffusion Transformer (DiT) backbones by changing only the cross-attention injection interface. On LIBERO-Mem, our model improves average success from 5% to 42.5%. On real robots, it achieves 78.3% success on counting tasks, 80.0% on spatial recall, and 75.0% on sequential tracking. Project page: https://MemoryVAM.github.io/

Tool-augmented large language model agents increasingly operate over large tool libraries, but existing evaluations often focus on whether a model can call a tool correctly rather than how the visible tool menu shapes reliability, efficiency, and safety-relevant risk exposure. We introduce ToolMenuBench, a benchmark for evaluating tool-menu construction in multi-step LLM agents. ToolMenuBench varies tool-menu size, distractor type, state-dependent task structure, and risk exposure, and reports both filter-level and downstream agent metrics, including visible-tool count, risky-tool exposure, task success, wrong-tool calls, premature actions, and token usage. In a controlled evaluation across seven model backends, three tool-menu sizes, six filtering methods, and seven evaluation settings, CMTF improves task success from 32.1% under all-tools exposure to 85.7%, while reducing average token usage by roughly 98%. Causal minimal tool filtering achieves the strongest overall tradeoff, reducing visible tools, wrong-tool calls, premature actions, and risky-tool exposure relative to unfiltered exposure, lexical filtering, state-aware filtering, and broader causal-path baselines. ToolMenuBench provides a reusable evaluation framework for studying the agent-interface problem: which tools should be visible, when they should be visible, and under what cost or risk constraints.

This exploratory observational study uses the Transluce platform to analyze the internal activity of Llama 3.1-8B-Instruct on 54 moral prompts in four batteries: dilemmas, policy, and meta-ethical questions (B1, 17); role-playing scenarios (B3, 6); and a structured trolley comparison varying the switching mechanism (B4, 15) or the people's identity attributes (B5, 16) with the other fixed. We record each surfaced neuron's normalized activation score, identity, and description and each cluster's title and size, classify them by an explicit description rule, and measure the prevalence and ranking of ethics-labeled neurons and clusters, the neurons' activation scores, and the model's verdicts and stated confidence. Prevalence is similar across batteries (about 5% of retained neuron rows) while ranking differs: in the trolley batteries the first ethics-labeled row has median position 3, versus 9 in B1 and B3, ethics labels are about four times as frequent among the ten highest-scoring rows as overall, and ethics-labeled rows have a higher mean activation score than other rows within prompts (47 of 54). We observe situational anchoring: prompt-setting vocabulary appears in prominent cluster titles and selected high-ranked neuron descriptions. B4-vs-B5 First Rank and Priority Ratio contrasts detect no difference. With one retained response per prompt, the coded outcomes for the 31 trolley responses are 19 recommending intervention, one recommending non-intervention, and 11 giving no definite recommendation. We distinguish prevalence, ranking, behavior, and causal contribution as audit targets, measure the first three, offer frame-conditioned moral computation as an organizing hypothesis with stated tests, and propose mechanistic alignment, causal responsibility of ethically relevant features under controlled frame variation, as a criterion admitting distributed implementations.

Large language models (LLMs) have facilitated impressive progress in software engineering, code generation, tooling, and systems. Concurrently, a significant body of research has developed which explores a growing variety of methods and systems for applying LLMs to hardware and chip design (e.g., systems for RTL code generation based on functional description). However, when it comes to open Verilog/RTL code-generation, we need high-quality training samples to build specialized and more effective LLM systems through fine-tuning or low-rank adaptation. Here, we propose a “judge-renew-check-renew-check” (JRCRC) pipeline which updates a current public dataset using a hierarchy of state-of-the-art commercial LLM models differing in their costs and capabilities in RTL code generation. This approach achieves a cost-effective mechanism for filtering and refining code-generation samples into a higher-quality training dataset. Our experiments also identify some common weaknesses of LLMs in rule-based reasoning and logic, and consequently, in RTL code-generation. Having identified these weaknesses, we develop an architecture for incorporating pre-processing tools to dynamically assist the LLMs in inferring logical relationships from tabular data formats. With our tools-assisted architecture for RTL code generation, we achieve significant overall performance gains in the VerilogEval benchmark and outperform many state-of-the-art methods. Our LLM4RTL system achieves performance comparable to that of GPT-4O using a significantly much smaller LLM.

Agentic AI systems act autonomously, use tools, adapt to context, and operate in complex real-world environments. However, these same characteristics can create or exacerbate product risks. We studied how industry developers (n=35) perceive, prioritize, and address the risks in their agentic AI products. We found that developers' perceptions of risk were closely tied to the qualities that made the product agentic, such as autonomy, tool use, and usage in a real-world context. Developers prioritized product and business risks before considering downstream societal risks like job displacement and end-user privacy. This prioritization also impacted developers' ability and motivation to mitigate agentic risks. Finally, developers lacked mature controls for containing agentic risks, often relying on constraining the same characteristics that make agents useful: e.g., autonomy and goal complexity. These findings reveal a capability vs. risk control tension in agentic AI development: developers need to address risks that emerge from agentic capabilities, yet they currently have limited support for doing so without constraining agentic functionality.

Steerable convolutional neural networks (Steerable-CNNs) guarantee SE(3)-equivariance by parameterizing kernels as linear combinations of steerable basis functions, but their deterministic nature precludes uncertainty quantification - limiting their use in settings where confidence estimates are essential. We propose a Bayesian Steerable-CNN that places posterior distributions over the basis coefficients, yielding stochastic kernels while preserving equivariance exactly. The loss function of the model is obtained via variational inference and minimized by Bayes-by-Backpropagation. The framework admits a decomposition of predictive uncertainty into epistemic and aleatoric components. Empirically, the model attains competitive classification accuracy alongside an expected calibration error of 0.0263 and outperforms its deterministic counterpart by up to 6.17% under distributional shift induced by additive Gaussian noise. Furthermore, we leverage the model's uncertainty estimates to enhance its performance significantly, achieving a notable gain - approximately 4% higher accuracy across 84% of the test dataset. A statistically significant negative correlation between epistemic uncertainty and prediction error confirms that the learned posterior variance is semantically meaningful. The framework unifies Bayesian uncertainty quantification with the inductive bias of equivariant CNNs.

Continuous evaluation of LLM products relies on a strong LLM judge treated as ground truth: a cheap monitor scores every interaction and a team is paged when the score drifts down. But the judge is itself a model behind an API, and a silent version bump or scoring-prompt update changes how it scores – so every drift alarm is ambiguous between a worse product and a changed judge. We resolve the ambiguity with a fixed, human-labeled anchor set that the current judge re-scores at a steady interleave, a second betting e-process on the judge-versus-human gap, and a guard-window rule returning a verdict in {none, system, judge}. We prove anytime-validity, one-way identification (only the judge can move the anchors), an attribution race whose design law is that the anchors must out-run the main process they guard, and process orthogonality. On two real judge changes, a silent version bump is detected as judge drift in 60/60 runs with zero judge-to-system misattribution, and a contaminating strict-prompt change is correctly attributed on 110 of 120 runs at guard width 300 – while the industry-default rolling z-test false-alarms on 75% of drift-free streams. Every experiment replicates on a second domain (TL;DR summarization) with nothing re-tuned, and where the domains differ the differences are the ones the race predicts: the strict-prompt change shifts scores harder there, so the anchors fire faster and attribution becomes perfect (240/240). The monitor runs at approximately 0.64 of the cost of strong-judging every item, or 0.21 in a cheaper-but-deafer regime.

Reinforcement learning with verifiable rewards (RLVR) has become a key approach for enhancing the reasoning abilities of large language models. However, RLVR often suffers from diversity collapse: Pass@$1$ improves while high-$k$ Pass@$k$ degrades, which is viewed as a narrowing of the model's reasoning boundary. We formalize this diversity collapse through the lens of overtraining: once a problem's contribution to the reference metric has effectively saturated, further updates no longer expand what the model can solve but still concentrate probability mass on the trajectories favored by on-policy sampling. Under a standard setup with few rollouts per problem, even a single observed success places a problem in a nearly saturated regime for high-$k$ Pass@$k$, so most updates in standard RLVR are overtraining from the boundary perspective. This perspective also suggests a reading of whether RLVR can expand the model's reasoning abilities beyond the base model: since RLVR is structurally biased against high-$k$ Pass@$k$, its aggregate decline does not by itself mean that no new reasoning gains occurred. Interventionally, restricting updates to problems with zero observed success lifts Pass@$256$ above the base model on difficult benchmarks; observationally, a non-trivial fraction of initially unsolvable problems become solvable during standard RLVR training. Building on these findings, we propose Bayesian Boundary Gating (BBG), which redirects optimization away from overtraining by estimating each problem's marginal contribution to the reasoning boundary. Across multiple reasoning benchmarks, BBG improves average Pass@$k$ across a wide range of $k$.

Electronic health record foundation models typically treat ICD diagnosis codes as flat tokens, overlooking the clinically meaningful hierarchical structure that captures disease families, subcategories, and fine-grained diagnostic detail. As a result, existing EHR representation learning methods do not explicitly exploit the hierarchical structure already present in the coding system. In this work, we study ICD-10-CM hierarchy as a general inductive bias for clinical representation learning. We investigate two complementary mechanisms for incorporating hierarchy: first, by augmenting diagnosis sequences in a BERT-style transformer with tokens corresponding to different levels of the ICD hierarchy, and second, by injecting hierarchy into graph-based code representations through hierarchy-aware edges combined with diagnosis co-occurrence structure. Across these settings, we evaluate whether explicit hierarchy improves downstream prediction, which levels of the hierarchy are most useful, whether hierarchy encoding improves transfer across datasets, and how hierarchy reshapes embedding similarity structure. We conduct experiments on two large-scale real-world clinical datasets: MIMIC-IV, used for pretraining and in-domain evaluation, and eICU, used to assess cross-dataset transfer via frozen encoder probing. Our findings show that explicitly encoding ICD hierarchy improves over flat code representations in both in-domain and cross-dataset settings, while revealing that the most useful level of hierarchy depends on both the task and the modeling approach. More broadly, we focus on hierarchy-aware EHR representation learning and show that the benefits of encoding hierarchy are generalizable across modeling settings and hierarchy levels.

Indirect prompt injection attacks hijack LLM-based agents by embedding malicious instructions in third-party data that the agent retrieves during task execution. Existing defenses report near-zero attack success rate on static benchmarks, yet recent adaptive evaluations show that these results collapse once the attacker is allowed to optimize against the deployed defense. In this work, we trace this collapse to two failure modes. First, existing defense methods are confined to recognizing specific attack patterns, rather than assessing whether the intent of every embedded instruction is relevant to the user task. Second, training-based defenses, which otherwise offer the strongest safety-utility trade-off, assemble their adversarial examples from a handful of hand-crafted templates, and the resulting defender fails to generalize outside that narrow strategy distribution. To address these gaps, we propose RETA, a training-based method that grounds defense decisions on the user tasks rather than attacker-controlled data. At each tool-output step, the defender undertakes chain-of-thought reasoning verifying that its actions are consistent with the user task. Leveraging red-teaming, a simulated attacker synthesizes adversarial training data and receives a dictionary-learning diversity reward, achieving broad coverage of injection-reformulation strategies. Together, these allow the defender to be optimized via multi-objective reinforcement learning and achieve better safety-utility trade-off. Across six black-box adaptive attacks, RETA keeps every per-attack ASR below 10%, with average ASR of 2.92% and 3.75% on the two target models, while preserving most utility under attack and on clean inputs.

Respiratory acoustic foundation models (FMs) excel at cough classification, yet their ability to predict continuous health quantities from cough audio remains largely unexplored, despite the clinical value of passive age, BMI, and disease probability estimation in settings where physical measurements are unavailable. We introduce the multi-model, multi-target cough regression benchmark evaluating five FMs (OPERA-CT, OPERA-CE, OPERA-GT, HeAR, M2D+Resp) across six targets on three datasets under subject-disjoint protocols, comparing linear, MLP-small, and full MLP regression heads. MLP-small beats the mean-predictor baseline on all tasks and linear probing in 23 of 30 model x task cases, with full MLP overfitting on small clinical data but recovering on larger sets, revealing a dataset size x head-capacity trade-off. HeAR leads within-dataset age regression on Coswara (9.12 yr MAE); its CIDRZ result is excluded from headline claims owing to possible HeAR-CIDRZ pretraining overlap. OPERA-GT is favored over OPERA-CT on age in all three datasets, with the CIDRZ margin within seed variance, extending a generative-pretraining advantage from breath to cough. HeAR and M2D+Resp reach near-full performance at N = 50 samples while OPERA models require N = 400. Cross-dataset transfer is strongly asymmetric as large diverse data generalises to small clinical populations (CoughVID to CIDRZ: -0.17 yr) but not vice versa (CIDRZ to Coswara: +2.43 yr, +26.6%).

Spaceborne inspection systems often deploy perception models prior to launch, after which updating model weights or expanding fixed label sets becomes operationally impractical. While supervised models can be integrated pre-flight, adding new semantic capabilities in orbit requires retraining and re-uploading parameters. We investigate whether prompt-driven vision–language models can enable post-launch semantic expansion, allowing new spacecraft components to be specified via natural-language prompts without modifying onboard weights. We evaluate zero-shot instance segmentation of spacecraft components under a strictly frozen, single-pass inference protocol on a test set of $129$ images of previously unseen satellites. Under fixed global thresholds and no post-processing, SAM3 achieves $0.385$ mAP@$0.5$ and $0.267$ mAP@$0.5{:}0.95$. Performance is strongly scale-dependent: large structural elements like spacecraft bodies ($0.639$ AP@$0.50$) and solar arrays ($0.598$ AP@$0.5$) localize reliably, while relatively small appendages like antennas ($0.221$ AP@$0.5$) and thrusters ($0.081$ AP@$0.5$) remain difficult. Prompt formulation influences performance, with structured prompts incorporating spatial and geometric descriptors yielding up to $82%$ improvement over short category-name prompts. The model operates within the memory and compute envelope of contemporary embedded GPUs, suggesting prompt-driven grounding can provide a practical mechanism for post-launch semantic extension of dominant spacecraft structures while highlighting limitations of zero-shot localization for fine-scale components under orbital domain shift.

Safety evaluations test a policy on prompts that omit the incentive information deployment supplies: a commission, a performance score, a dashboard naming which action pays best. We measure what that omission hides. In MoneyWorld, a synthetic workplace environment, we train five instruction-tuned models from three families with RL on non-safety tasks in which a visible payoff signal identifies a rewarded shortcut that sacrifices task quality. We then freeze each policy, present held-out safety conflicts, and change only the displayed signal. Each menu contains one compliant action and three violations. We report three findings, with rates for Qwen2.5-14B-Instruct. (i) Payoff signals control frozen safety choices: unsafe choice is 100% when the signal names an unsafe option and 0% when it is hidden or names the safe one. Hidden- and random-signal training controls stay at or below 0.3%, and the switch reproduces on all five bases. Numerical payouts reproduce it under sampled-action rewards, reaching 98.6% unsafe choice at a $1 advantage. (ii) Payoff identification and unsafe choice separate under a training-menu intervention: training on task-completing actions at the same payouts retains 99.8% identification while reducing unsafe choice to 7.9% at matched update budgets. Payoff-reading competence alone does not explain transfer. (iii) The switch does not reproduce in executed retail customer-service tasks using the same frozen adapters. In MoneyWorld, omitting incentive information conceals unsafe choices that appear when the same policy sees which action pays best.

Objective: To enhance the accuracy, interpretability, and robustness of large language models (LLMs) in medical question answering (MedQA). Method: We designed a multi-agent peer-reviewed reasoning method in which multiple LLM agents independently generate chain-of-thought reasoning with candidate answers, then act as peer reviewers to evaluate each other's reasoning for factual correctness and logical soundness. The highest-rated reasoning chain is selected to produce the final answer. Experiments were conducted with five state-of-the-art LLMs (Llama-3.1-8B, Qwen2.5-7B, Phi-4, DeepSeek-LLM-7B, GPT-oss-20B) on three benchmark datasets: HeadQA, MedQA-USMLE, and PubMedQA. Performance was compared against single-model chain-of-thought reasoning and chain-of-thought-based majority voting. Results: Peer-reviewed reasoning consistently outperformed both baselines. The best model combination achieved an average accuracy of 0.820 across datasets, exceeding the strongest single model (0.777) and majority voting ensembles (up to 0.789). The method also scaled effectively with more participating models, while peer assessments reliably distinguished high- from low-quality reasoning chains. Conclusion: The proposed multi-agent peer-reviewed reasoning method enables LLMs to act as both solvers and evaluators, yielding superior performance in MedQA. By emphasizing reasoning quality rather than answer agreement alone, this approach improves accuracy, interpretability, and robustness, offering a promising direction for trustworthy biomedical AI systems.

Biomedical relation extraction (BioRE) is a key step in transforming biomedical literature into structured knowledge. Most existing approaches rely on supervised models trained on costly annotated datasets, limiting their scalability and adaptability across relation types and domains. We investigate few-shot BioRE using prompt-based learning with large language models (LLMs) and compare two task formulations: pairwise classification, which predicts relations for individual entity pairs, and joint generation, which extracts multiple relations in a single model call. Experiments on the BioREDirect dataset reveal a clear precision-recall trade-off. Pairwise classification achieves higher recall, whereas joint generation is more precise and computationally efficient. The best-performing model achieves a micro-F1 score of 0.44, substantially outperforming previous few-shot results (0.34) while remaining below the supervised baseline (0.56). Much of this gap is attributable to a single ambiguously defined relation type. When evaluated using macro-F1, which better captures performance across relation types in an imbalanced setting, prompt-based approaches outperform the supervised baseline (0.45 vs. 0.38), particularly on rare relation types. These findings highlight the potential of LLMs for BioRE in low-resource settings and underscore the importance of well-defined relation schemas.