Papers

Evaluating the average packet error rate (PER) of finite-blocklength (FBL) coded transmission over fading requires integrating the block-error waterfall, given by the normal approximation, against the fading distribution, which is intractable for general Nakagami-$m$ channels. This letter approximates the conditional waterfall by a slope-matched logistic function and shows that its Nakagami-$m$ average reduces to a single Lerch-transcendent term that interpolates between the FBL waterfall and the classical outage limit, with norming constants explicit in rate and blocklength. The closed form supports non-integer fading and, composed into an effective-capacity objective, yields a quality-of-service (QoS) aware rate-selection rule. It matches the normal-approximation integral to about 1% uniformly in $m$ over the nominal FBL operating region, while outage and linearization baselines exceed several percent at high diversity.

We show that border rank is equal to rank for Kruskal tensors. We also give an analogous Kruskal condition for alternating tensors, which certifies uniqueness of skew rank decompositions. Furthermore, we show that border skew rank is equal to skew rank for alternating Kruskal tensors, and we give an algorithm to find the minimum skew rank decomposition of alternating Kruskal tensors.

Fisher width measures the Gaussian width of a probe set after deformation by the local Fisher geometry. We study its evolution along learning trajectories and ask when training loss can serve as an effective coordinate for this quantity. We first derive an exact trace–shape factorization and a deterministic stability bound for fixed compact probes. In a population Gaussian-teacher logistic model, the teacher-aligned state is extremal on every loss level below $\log 2$: it has minimal parameter norm and maximizes both Fisher trace and Euclidean-ball Fisher width. We then show that population gradient flow asymptotically selects this branch, with explicit rates for the aligned and orthogonal coordinates. This yields, for $d\geq2$, \[ \frac{w_F(B_2^d;\theta(t))} {\sqrt{L(\theta(t))}} \longrightarrow \frac{\sqrt6}{\pi}\mathbb E[\chi_{d-1}]. \] Controlled full-Fisher experiments support the matched-loss branch and the population predictions. In a nonlinear MLP with a diagonal model-Fisher approximation, GD and SGD remain close at matched loss, whereas Adam follows a substantially displaced branch; the fixed probes tested retain highly similar temporal shapes. These results support a branchwise, rather than universal, loss parametrization of Fisher width.

Multi-stage LLM pipelines can remain structurally valid even when evidence available to downstream stages becomes incomplete, compressed, or conflicting. This paper introduces and operationalizes Evidence-State Reliability (ESR), an evaluation layer concerned with whether intermediate evidence remains sufficiently complete, grounded, internally consistent, and usable for a stage's assigned function. ESR is evaluated separately from parser validity, which measures structural conformance. We evaluate the framework using GLM-5.2 on 60 sanitized base cases under four evidence conditions: clean, compressed-lossy, partial-dropout, and noisy-conflicting. Each condition was processed through decision, audit, and escalation stages. The design comprised 720 planned and ledgered calls, with 713 retained, sanitized execution rows. Across nine matched degraded-minus-clean condition-stage comparisons, all operational stage-success estimates were negative, and all 95% bootstrap intervals remained below zero. All nine parser-validity point estimates were positive, although the three partial-dropout intervals included zero. Among parser-valid degraded audit outputs, degradation detection was 1.0 in each degraded condition, while false-assurance rates remained non-zero; among parser-valid degraded escalation outputs, recovery was 0.0 in every degraded condition. The results show a bounded reliability-layer divergence in the evaluated pipeline: structural conformance can improve directionally while evidence-sensitive stage success deteriorates under the same controlled intervention. They also separate detection of degraded evidence from recovery. The conclusions are limited to the evaluated model configuration, pipeline design, selected sanitized cases, scoring procedure, and single scaled run.

Recent advances in LLMs and the adoption of RAG systems in industry have created a need for domain-specific question-answer datasets that can assess RAG performance on proprietary data. Existing datasets, such as HotpotQA, challenge current RAG systems on Wikipedia-based knowledge, but they cannot be transferred directly to domain-specific settings. A comprehensive evaluation of RAG system quality requires both multi-hop queries and unanswerable questions. This paper introduces TRIAD, a three-stage automated dataset generation approach. First, it generates question–answer (QA) pairs for the domain-specific knowledge base of a RAG system. Second, a validator checks each QA-pair in a feedback loop. Third, the QA pairs are extended with relevance-labeled context documents for downstream evaluation. We evaluate this approach against the established MuSiQue and HotpotQA datasets. The results show that the generated dataset exhibits similar performance trends across different RAG setups, while human validation indicates that the questions are suitable for evaluating a domain-specific RAG system. The code used to generate the dataset and all validation results are available in our GitHub repository(https://github.com/lorenzbrehme/triad).

Modern machine learning compilers select tensor memory layouts to minimize execution cost under hardware constraints. Layout selection is global: an operator may be fastest under one layout while its consumers prefer another, and aligning these preferences requires explicit layout conversions that can hurt model performance. Despite its practical importance, layout selection lacks a formal basis, so current compilers rely on ad-hoc heuristics. This paper presents the first formal study of layout selection in machine learning compilers. We formulate the problem as combinatorial optimization over dataflow graphs, minimizing the sum of operator execution costs and the per-tensor cost of these conversions. Our theoretical analysis shows that optimal layout selection is computationally hard, even for programs containing only matrix multiplications over two-dimensional tensors. We design an optimal polynomial-time algorithm for dataflow graphs of bounded treewidth. For general instances, we give a weighted MaxSAT encoding that an off-the-shelf solver can optimize. The formulation unifies several existing layout optimization strategies, including XLA's layout assignment, partition dimension selection in systolic array compilers, and layout planning in mobile GPU optimizers. We implement the formalization in a production compiler for an AI accelerator and measure the execution time of the compiled models under greedy heuristics, the compiler's rule-based strategy, and an optimal solver. Simple heuristics degrade execution time by up to $5\times$ on some workloads. Where the compiler's cost model is accurate, the solver matches or beats the rule-based strategy. On workloads with complex data movement it falls behind, and since the solver minimizes the stated objective exactly, that gap isolates cost-model error from search quality, showing where compiler effort actually pays off.

Heterogeneous multi-robot service systems must assign requests to compatible robots, construct feasible schedules, and adapt as new tasks arrive online. Historical data can help anticipate future demand, but relying too heavily on inaccurate predictions can degrade performance under distribution shifts. We develop a prediction-aware adaptive rollout framework for heterogeneous multi-robot task assignment with scheduled and real-time requests. The problem is formulated as a finite-horizon stochastic dynamic program incorporating robot-task compatibility, ordered service requirements, routing constraints, service windows, and end-of-horizon return requirements. The proposed policy evaluates current assignments using sampled future request scenarios while restricting immediate commitments to requests already observed. To enable online use, the framework combines pruned candidate controls, wait actions, and an interaction-aware base policy for efficient future-cost estimation. Robustness to forecast error is provided by adaptively reweighting predicted requests based on recent prediction mismatch and selectively re-optimizing assigned but unstarted requests. We also introduce a historical-data-driven procedure for selecting the heterogeneous fleet composition before deployment. In a case study using real nursing-task requests from hospital inpatient floors, the proposed approach achieves near-complete service and reduces serviced-request wait times relative to reactive, token-passing, prediction-positioning, and myopic greedy baselines, with the largest improvements in tail-delay metrics.

Complex-valued arrays arise in signal processing, where scientific interpretation depends on retaining amplitude and phase information. Existing covariance estimation methods either ignore the multiway organization of such data or rely on real-domain embeddings that do not directly exploit their complex structure. We develop sparse separable factor analysis (SSFA), a latent factor model for complex-valued arrays with a separable covariance structure across modes. Each mode-specific covariance matrix is modeled through a low-rank Hermitian factor structure and a diagonal residual covariance matrix. To obtain interpretable estimates, we impose elementwise lasso penalties on the complex loading matrices and estimate the SSFA parameters using a mode-wise parameter-expanded expectation-maximization procedure. The resulting loading updates admit closed-form complex soft-thresholding solutions, which shrink the modulus of each loading while preserving its phase. A separate balancing step resolves the scale nonidentifiability of the separable covariance structure. Simulation studies show that SSFA improves covariance estimation relative to vectorization-based methods, including complex principal component analysis. We apply SSFA to local field potential recordings from mice, where we compare separability structures induced by different groupings of brain region, frequency, and time and perform model-based imputation of recordings missing because of electrode misplacement.

Disassembling end-of-life electric vehicle (EV) battery packs is dull and dangerous work, performed almost entirely by humans. We present GOLEM (Generalized Open Library of Embodied Modules), an end-to-end, open-source system architecture for EV battery disassembly with the Unitree H1-2 humanoid robot in which walking, manipulation, dynamic stability, navigation, and spatial memory are independent modules with abstract interfaces, so that methods are easily developed, interchanged, and compared. GOLEM is deployed as a Docker-based ROS 2 abstraction in which MuJoCo and IsaacLab digital twins expose interfaces matching the physical robot. GOLEM's composability and per-module customization enable development and demonstration of humanoid EV battery disassembly, from simulation to reality. GOLEM provides fair comparison between humanoid modules, enabling evaluation as a capability ladder, in which one module is characterized at a time and added as a rung: LiDAR-inertial navigation places the robot within 13.0cm of a 6m goal; a learned standing controller recovers from external disturbances that sampling-based lower-body MPC does not; and grasping loosened fasteners from a real Hyundai Ioniq 5 pack degrades from 97% tethered to 87% free-standing to 37% under navigation-induced pose variance. Source code is available at the project page https://golem-humanoid.github.io

The rapid development of artificial intelligence, including agents and deepfake techniques, has accelerated phishing attacks and lowered the threshold for attackers. Modern phishing attacks now blend multiple tactics, including social engineering, URL spoofing, and AI deepfakes enabling adversaries to craft highly convincing messages that exploit human vulnerabilities and bypass traditional detection systems. At the same time, current security awareness education struggles to keep up with the speed, sophistication, and complexity of these evolving threats. To address this challenge, we propose a two-stage anti-phishing framework, CyberGLA, that combines technical defense and user-centered security education. In the Detection stage, we introduce EmailKnight, a spoof detection tool that performs multi-level email analysis. To enhance user awareness, the Training stage incorporates a large language model (LLM)-based security coach that dynamically selects personalized training modules based on the outcomes of the Detection stage. This dual purpose design philosophy enables effective protection against the evolving threats of modern email phishing attacks.

The primary challenge of conducting PDE-constrained optimization for high-dimensional problems, such as kinetic equations, is the often prohibitive memory cost. Computing gradients using the adjoint state method would require the storage of the entire time history of the forward solution. For such problems, where the memory cost for storing a single instance of the forward solution can already be a limiting factor, this is clearly not feasible. In this paper, we propose a memory-efficient adjoint state method that compresses the forward and adjoint solution with a dynamical low-rank approximation (a model order reduction technique) and bypasses the need to store the entire forward solution by employing a time-reversible low-rank integrator. The dynamical low-rank approach introduces a number of challenges: reversibility can fail in the rank-deficient case and the low-rank trajectories can show chaotic behavior. In particular, the latter has a number of important consequences for the optimization problem. We address those challenges and show that our method can drastically reduce the memory requirement for gradient-based optimization of kinetic equations. In particular, we consider two examples from kinetic plasma physics: optimizing beam profiles to suppress a bump-on-tail instability and shaping a particle beam using external electric fields.

Large language models (LLMs) are increasingly deployed as tool-augmented agents, where responses can depend on tool calls and external observations rather than model parameters alone. This creates an evaluation mismatch for LLM unlearning: previous unlearning methods may suppress direct parametric recall, but an agent can still recover the same forget target through tools such as web search, retrieval, or database lookup. We identify this failure mode as tool-mediated recovery and study agentic tool unlearning, which aims to reduce both parametric recall and tool-mediated recovery while preserving normal tool use for retained knowledge. To address this challenge, we propose Agentic Tool Unlearning (ATU), a two-stage framework. The first stage applies parametric knowledge unlearning to suppress direct recall, while the second stage performs trajectory-level reinforcement learning in simulated tool-augmented environments to penalize target-seeking tool behavior and final-answer leakage. Experiments on RWKU and MUSE across different LLM architectures show that ATU achieves a better balance between target forgetting and retained utility, making unlearning more robust under tool-augmented agent deployment.

Object placement is critical in image composition, requiring spatially and semantically coherent positioning of objects within diverse scenes. Existing approaches typically rely on hand-crafted rules or supervised learning on limited datasets, which restricts their generalization and interpretability, especially in open-world scenarios involving novel objects and scenes. In this work, we reformulate open-world object placement as a heuristic search task guided by reasoning from a Multimodal Large Language Model (MLLM). We introduce presto, a zero-shot, training-free framework that operates within an imaginary action space to iteratively refine object position and scale. Our coarse-to-fine search strategy ensures fast convergence, and we evaluate two decision-making variants: Metric-guided Selection and MLLM-as-a-judge. Experiments across multiple benchmarks show that presto~achieves state-of-the-art performance, particularly in previously unseen, open-world settings. Human studies further reveal that the MLLM-as-a-judge variant produces more perceptually coherent placements than metric-driven approaches, highlighting a gap between standard evaluation metrics and human visual judgment.

KV-cache compression is often justified by attention maps with a few large weights. This is incomplete: large weights may not contain most of the mass, omitted values can cancel, and preserving the attention output may not preserve the task. We separate these questions. Global score gaps – not threshold counts – determine how many tokens are needed to retain a target mass. For a realized row, the weighted sum of omitted values is the exact missing statistic. A controlled retrieval–aggregation model explains when truncation helps and when it hurts. These results motivate CertKV, a training-free compressor that reserves one tail-summary slot per head and allocates the rest by value dispersion. Under matched budgets, CertKV is top-two in seven of nine LongBench-v2 settings, remains in the leading compressed tier on 128K RULER, and realizes a ten-fold cache budget in a packed Llama prototype. Compressibility depends on the mass, values, future queries, and task – not on a sparse-looking map alone.

Federated Continual Learning (FCL) is fundamental to real-world distributed learning systems, requiring models to adapt to sequential, non-IID data across clients while mitigating catastrophic forgetting and client drift. Existing approaches formulate continual learning (CL) as a sequence of per-task optimization problems, applied locally at each client and coupled through aggregation, using heuristic mechanisms such as replay, regularization, or projection-based constraints. However, forgetting in FCL is inherently a long-term, distributed phenomenon, arising from the interaction of temporal task evolution and cross-client heterogeneity, which is not explicitly regulated. In this work, we cast FCL as a stochastic control problem and propose Federated Queue-regulated Continual Learning (FedQCL), a framework based on Lyapunov drift-plus-penalty (DPP) optimization. FedQCL introduces virtual queues to track the accumulation of forgetting across tasks and clients, enabling explicit control of the stability-plasticity trade-off. By optimizing a DPP objective, the method jointly improves current-task performance while the queue-based formulation provides an interpretable and tunable mechanism to balance adaptation and retention through a single parameter, without requiring gradient projection or additional communication overhead. Empirical evaluations on standard benchmarks, including Split-CIFAR-10, Split-CIFAR-100, and Split-TinyImageNet, demonstrate that FedQCL outperforms state-of-the-art baselines with respect to accuracy while significantly reducing forgetting under heterogeneous data distributions.

Automated applicant tracking systems increasingly decide who advances in hiring, and litigation and regulation now demand that those decisions be auditable. Existing tools sit at two extremes. Group fairness metrics such as the disparate impact ratio summarize a whole population but cannot say which individual decisions were unfair or why, while local explainers such as SHAP attribute a single prediction but are not connected to the legal standard by which hiring bias is judged. We present the AI Bias Firewall (AIBF), a method that audits an applicant tracking system one decision at a time. AIBF neutralizes a candidate's protected-attribute proxies, re-scores the decision, and measures the resulting counterfactual shift, which yields a signed per-decision bias in score points, a flag for decisions the protected attributes changed, and a plain-language explanation naming the responsible factors. We evaluate on two real public datasets, Adult and COMPAS, rather than on synthetic data. The per-decision counterfactual shift is faithful, aggregating to reproduce the known group level disparity, for example a mean shift of +7.5 points for the privileged group and -8.0 for the disadvantaged group on Adult, consistent with the measured statistical parity difference. AIBF identifies the decisions that protected attributes flipped with an area under the ROC curve of 0.963 on Adult, against 0.672 for a baseline that flags by group membership, and it identifies the harmed candidates so precisely that reviewing only five percent of decisions surfaces fifty-five percent of them, against six percent under group based review. We also report a limitation: correcting flagged decisions raises the disparate impact ratio substantially but not to legal parity, because features labeled as merit carry residual proxy correlation. AIBF is released under the Apache 2.0 license with code and experiments.

Modern hardware designs for AI and cryptography treat data transit and processing separately. At Optalysys we have demonstrated programmable Photonic hardware that computes mathematical functions on data that is in transit, enabling tens of GFLOPs of operations on data links.

Robotic Fulfillment Centers (FCs) store inventory on shelves (pods) arranged in dense blocks. Retrieving a target pod that is buried deep in a block requires moving obstructing pods out of the way (i.e., digout). Multi-robot planners use parameterized cost functions to control digout behavior, producing a spectrum of strategies: at one extreme, obstructing pods are sent to other blocks (using more robots in travel lanes); at the other, pods are shuffled within the block (avoiding lane congestion but increasing extraction time). Each point on this spectrum has different downstream consequences for floor congestion and throughput. The optimal operating point depends on the specific facility configuration and shifts with operational conditions such as varying station demand and congestion patterns, making offline tuning impractical. We present an adaptive parameter tuning framework based on Extremum Seeking Control (ESC) that continuously adjusts planner parameters in response to measured throughput. ESC performs model-free optimization by perturbing parameters with sinusoidal dither signals and correlating perturbations with performance changes to estimate gradients, making it robust to the multi-minute delayed effects and credit assignment challenges inherent in large FC operations. Simulation studies demonstrate that the adaptive policy improves upon fixed policies across several conditions. We observe an improvement in throughput by an average of 5.0% across map and robot fleet size variations, and by 8.4% under dynamic operating conditions. This work eliminates manual parameter provisioning and enables real-time adaptation, providing a self-tuning paradigm for FC storage operations.

Requirements-related information is scattered across heterogeneous artefacts such as user feedback, developer discussions, and software repositories, making the extraction of actionable requirements knowledge labour-intensive and hard to scale. Large Language Models (LLMs) can support many Requirements Engineering (RE) activities, from classification and traceability identification to specification and explanation generation, but existing evidence is fragmented across tasks, artefact types, and evaluation settings, and studies rarely offer cross-task evaluations or replication packages. We present two complementary empirical studies evaluating LLMs across five RE-related activities. The first is a controlled experiment on five lightweight open-source LLMs for feedback-driven requirements classification and specification generation. The second is an exploratory industrial case study on two frontier LLMs for traceability link identification and traceability explanation generation using real project artefacts. Classification and traceability identification were assessed with quantitative metrics, and generation tasks through human evaluation. LLM performance is strongly task-dependent, ranging from moderate to high, and no single model consistently outperformed the others, indicating that effective adoption depends on selecting models and prompting strategies per task. Our contributions are: (i) the first cross-task empirical evaluation of LLMs spanning five RE-related activities, (ii) replication materials supporting reproducibility, and (iii) a broader understanding of the capabilities, limitations, and practical readiness of current LLMs for RE.

Machine learning has had a positive impact on the sports industry, with one of its most promising applications being the prediction of athlete performance and injury risk. Recent advances have employed state-of-the-art models to improve prediction accuracy, yet progress remains limited by data availability and the reliance on subjective observations or expert assessments. To address these limitations, researchers in sports such as soccer, basketball, and wrestling have begun integrating heterogeneous data sources, such as wearable device readings, with traditional subjective assessments. However, similar multimodal approaches remain underexplored in tennis. In this work, we propose a multimodal weighted ensemble learning framework, Predictive Athlete Readiness for Tennis (PART), to monitor athlete wellness and estimate near-term injury risk in tennis players. PART processes a wide range of inputs, including physiological metrics, training and match data, sleep information from wearable devices, self-reported questionnaires, vertical jump assessments, and motion analysis from match-play videos. From these modalities, specialized machine learning and deep learning models independently extract four athlete-specific characteristics: overall wellness, injury risk, physical capability, and playing style. To overcome the complexity of combining these diverse modalities, PART employs a supervised weighted ensemble integration strategy, assigning adaptive weights to each predictive model based on its reliability. Evaluation of multimodal data collected from nine collegiate tennis players demonstrates that PART achieves strong performance in monitoring athlete wellness and estimating near-term injury susceptibility. Beyond collegiate athletes, the framework also shows promise for recreational tennis players, offering personalized insights to mitigate injury risk and optimize performance.

Timely and accurate assessment of property damage is critical following natural disasters. Traditional on-site inspections are labor-intensive, costly, and often pose safety risks. Advances in satellite imagery and vision-language models (VLMs) enable scalable remote damage assessment; however, integrating VLMs into large-scale Earth observation pipelines presents challenges in computational efficiency, data organization, and information retrieval. To address these challenges, we present DamageScope, a retrieval-augmented framework that combines satellite imagery with Vision-Language Models (VLMs) and Large Language Models (LLMs) to automate property damage analysis. Built on a Retrieval-Augmented Generation (RAG) framework, DamageScope extracts structured visual representations from satellite imagery to support interactive natural language queries for damage assessment. To address scalability, we introduce a novel multi-vector embedding-based clustering algorithm that outperforms traditional single-vector embedding approaches while reducing indexing time by up to 14x. Furthermore, a dual-store data architecture minimizes LLM API calls, reducing both operational cost and response latency by up to approximately 3x. By effectively balancing scalability and operational efficiency, DamageScope provides a robust and practical solution for real-world damage assessment tasks.

One aim of neural network field theory is to put a quantum or effective field theory on a computer, with the network ensemble itself as the theory. We ask how far that aim can be pushed for a function class regular enough to be computed with. Our main result is a no-go theorem with assumptions that hold for standard network architectures. We use it to separate four versions of neural network field theory, according to whether the defining object is the finite width ensemble or its infinite width limit, and whether the target we want to compute is a quantum or an effective field theory. Neither finite width interpretation is straightforwardly consistent. For finite width ensembles with finite variance at each point, the QFT interpretation fails reflection positivity, while the EFT interpretation establishes no scale separation by which the positivity violation can be placed outside its domain of validity. Of the two limit versions, one can be simulated in full and the other only in part, as only its smeared correlators are computable with a controlled error. As such, at the level of a controlled numerical computation, the QFT and EFT versions cannot be distinguished. One dimension escapes the obstruction, yet reflection positivity is shown to still fail there at every finite width for the cosine network. Two escapes from the theorem remain, giving up either finite variance at a point or exact rotation invariance, and we discuss both of these possibilities.

Formal verification provides the strongest correctness guarantees for software, and verification-aware languages can produce sound, machine-checked proofs. Recent AI coding agents have sharply lowered the cost of constructing such proofs. Yet few mainstream developers benefit: most use languages without formal-verification support, and formalizing properties and modeling execution environments demand formal-methods expertise. Proof therefore remains reserved for a few notable artifacts, while production software is attested mainly through review and testing. We introduce neuro-formal verification (NFV), which brings this automation to mainstream languages. An AI coding agent formalizes a source-level verification problem into a proof obligation in a verification-aware language, discharged by an established sound verifier aided by agentic proof search. Staged, goal-blind transformations reduce the risk of proving an artifact that does not faithfully represent the source program, property, or environment. Since NFV cannot ensure the soundness of this formalization, it optimizes for empirical accuracy rather than end-to-end soundness, while insisting on machine-checked evidence for every verdict. Experiments with current frontier models on a balanced dataset of correct and buggy Python solutions demonstrate the effectiveness of our approach. NFV with Dafny correctly resolves 57% of all entries, at 92% precision among its verdicts; with a CBMC backend, it produces a counterexample for 63% of the buggy programs at 90% precision. In contrast, an LLM-as-judge baseline achieves only 72% precision while answering every entry without any checkable artifact, and an unstaged agent-verifier combination proves 98% of both the correct and the known-buggy programs, yielding only 50% precision. Together, they confirm that both proofs and staging benefit an AI agent's formal program reasoning.

The rapid advancement of large language models (LLMs) has led to increasing interest in their application to scientific domains such as chemistry. However, existing chemistry benchmarks often provide only a narrow view of model capability, focusing on limited task sets while overlooking robustness to variations in problem formulation and chemical representation. As a result, reported performance may overestimate a model's true ability to reason consistently across realistic settings. To address this challenge, we introduce ChemDIRT (Diversified Instruction, Representation, and Task Benchmark), a comprehensive evaluation framework designed to assess the robustness of chemical reasoning in LLMs. ChemDIRT systematically measures model performance across variations in instructions and molecular representations while spanning eight categories of chemistry tasks. By evaluating both accuracy and consistency under these controlled perturbations, ChemDIRT provides a more reliable assessment of model reasoning capabilities than conventional single-format benchmarks. We benchmark a diverse set of open- and closed-source LLMs, revealing substantial prompt sensitivity, representation dependence, and uneven performance across task families.

A verification substrate is more credible when exposing errors in its own claims, not just model outputs. GBI-DCSE v3 falsified an architectural claim: the reported Fisher value epsilon ~ 0.066 satisfies the kappa^2 <= 10^4 budget only on the slice [epsilon, 3, 4, 5], while the full box [epsilon, 20]^4 requires epsilon ~ 0.326472. This erratum highlights whether an enterprise verification architecture can isolate interface failure, task competence, policy admissibility, and control integrity while keeping claims auditable. BoundaryBench v0.1 established the baseline: Qwen3-4B-Instruct-2507 completed 768 frozen executions, but 0% cleared the contract (369 failed parsing, 399 failed validation), limiting downstream selectivity metrics. This companion study evaluates three successive improvements. First, BeTaL-GBI v0.2 applies Benchmark Tuning with an LLM-in-the-loop over 2,218,750,380 grid points, separating format admission from conditional performance (rho_adm = N_admitted/N; rho_task = N_verified/N_admitted). Following schema repair, a model-free feedback search achieves a 2.87% mean held-out target gap, outperforming non-feedback baselines (13.61%, 11.46%). Second, GBI v2 swaps static keys for a reference-independent witness state W and policy P. Across 512 synthetic tasks, a 16-gate policy detects all 116 injected severe contradictions and accepts all 99 clean records (broad denominator: 4.27%). Hallucinator and evidence-forger surrogates are blocked with zero silent promotions. Third, GBI-DCSE v3 maps 99 claims to machine-readable evidence: 95 of 96 testable claims pass, with 148 standalone checks executed without failure. The harness exercises signed ledgers, PBFT quorums, and enclave forgery across 62 configurations. Under synthetic conditions, GBI-DCSE is a selective, policy-versioned, self-auditing test and routing substrate.

Credit assignment in large-language-model reinforcement learning (LLM RL) can be separated into three objects: evidence about success, a transport operator that converts this evidence into token-level advantages, and an update geometry that turns advantages into policy changes. Recent work has greatly improved evidence, sampling, and update geometry, but the transport operator is usually architecture-agnostic. Fixed-discount GAE applies a stationary geometric kernel along token time; group-relative methods broadcast an outcome statistic across an entire response. Neither operator represents the trajectory-specific computation used by the Transformer policy itself. We introduce computation-conditioned credit transport (CCT), a general framework in which a detached statistic of the behavior policy's internal computation parameterizes the causal kernel that transports downstream value through a rollout. Our concrete algorithm, CompPO, maps native attention concentration to a bounded per-token retention gate, uses the gate in both the one-step bootstrap and a path-dependent generalized-advantage trace (Comp-GAE), and co-designs a transport-aligned critic (TAC) that reuses the actor's hidden states and routing information without a second same-scale Transformer. The task reward and clipped PPO policy objective remain unchanged; a constant gate recovers fixed-coefficient GAE. Across five Qwen3-4B seeds, CompPO reaches 61.4% final held-out accuracy (95% CI [60.8,62.0]) versus 53.8% [52.9,54.7] for tuned GRPO. Neither Comp-GAE with a standard critic (55.2%) nor TAC with a fixed gate (56.4%) matches the full model (interaction +2.4 [1.9,2.9]). Shuffle and position controls confirm trajectory-specific alignment; CompPO is stable in 10/12 PPO-grid runs versus 3/12. Frozen evaluation improves over GRPO by 4.3 and 3.9 greedy pass@1 macro points on Qwen3-4B and Llama-3.1-8B-Instruct.

Prompt injection is listed as the \#1 threat to AI agents. When an agent accesses external data from websites, files, or emails, an attacker may inject a prompt into the data, saying, "Ignore all prior instructions and perform <an attacker's task>." To prevent arbitrary manipulation of agents, defenders try to train secure LLMs, which, however, still suffer from near 100% attack success rates (ASRs) against adaptive prompt injections. We note that this is because existing defensive finetuning recipes rely on sequence-level feedback signals (in DPO or GRPO). Treating an entire output equally prevents the model from learning precisely which output tokens are insecure. In this paper, we propose Secure On-Policy Distillation (SecOPD) that provides token-level feedback to guide defensive fine-tuning. The LLM receives an injected sample and produces a rollout, whose tokens are scored by the initialization model given the corresponding clean input. With more fine-grained training signals, our defended Qwen3.6-27B achieves a 9.0% ASR against the SoTA PISmith adaptive prompt injections, compared to 94.0% for the prior SoTA, Meta-SecAlign. The obtained security generalizes to domains completely unseen in training: in agentic tool calling, SecOPD achieves a 4.7% ASR compared to 5.5% for Meta-SecAlign. Code and the model are available at https://github.com/pppyb/SecOPD and https://huggingface.co/pybbb/Qwen3.6-27B-SecOPD.

Cardiac auscultation remains the most cost-effective screening procedure for cardiovascular diseases, and requires listening at the four main auscultation spots. Despite this, automatic heart sound analysis algorithms mostly classify patients using a single heart sound (single-channel), or, when using more than one (multi-channel), analyze each channel individually. To our knowledge, no prior work classifies patients through the synchronous analysis of multi-channel heart sounds, following the procedure used by physicians. This motivates us to study whether synchronous multi-channel analysis outperforms single-channel approaches, and whether it holds an advantage over asynchronous multi-channel methods that analyze channels one by one, potentially by capturing inter-channel interference phenomena. To answer these questions, we introduce a selection algorithm that identifies optimal heart sound segments from each of the four auscultation spots, which are then fed into a multi-input CNN that classifies patients by analyzing the four selected sounds simultaneously. Our synchronous approach, combining the proposed selection algorithm with a multi-input CNN, achieves a superior overall accuracy of 96.5%, a 9.1% gain over the best-performing single-channel and asynchronous multi-channel methods. The benefit of the proposed segment selection strategy over random selection is confirmed by a paired statistical significance test ($p = 0.003$). These results were obtained on 735 patients from the CirCor DigiScope dataset with complete recordings from all four spots, and their scope and generalizability are discussed in light of this and other methodological considerations.

Biochemical recurrence (BCR), defined as any detectable prostate-specific antigen level after prostatectomy with confirmatory elevation, is widely used as a surrogate endpoint and typically assessed using clinical and pathological variables. Currently, no standardized benchmark exists for multimodal prognostic modeling in urological cancers, partly because curating heterogeneous multimodal data remains challenging. We developed the CHIMERA Challenge, a multimodal benchmark integrating preoperative mpMRI, post-prostatectomy histopathology, patient characteristics, and clinician-derived variables from 267 patients across two institutions. The dataset comprises 801 MRI sequences, 13 clinical variables per case, and 942 WSIs. Training (n=95), validation (n=23), and test (n=149) splits were established and hosted on the Grand Challenge platform. Baseline clinical and pathological characteristics did not differ significantly across splits. Models were evaluated on predicting time to BCR using the C-index. Post-challenge analyses tested how each model type performed when clinician-derived variables were withheld or randomized. Unimodal clinical models achieved the highest test C-index of 0.7402 but proved sensitive to the integrity of these variables, with performance collapsing toward chance (C approximately 0.50) when they were randomized. Multimodal models retained near-baseline performance when these variables were withheld (delta C at most 0.04), indicating their ability to recover prognostic signal directly from imaging data. CHIMERA is the first public, standardized multimodal benchmark for prostate cancer prognosis. Although models using only patient characteristics and clinician-derived variables yielded the highest leaderboard performance, multimodal models demonstrated greater robustness in clinically realistic scenarios where complete expert annotation is not guaranteed.

AI systems increasingly search among candidate answers and deploy the highest-scoring one. Increasing search changes which errors matter, so a precise evaluation at one computation budget can leave another budget unresolved. We connect this information gap to the cost of closing it. For independent best-of-n search, aggregate reliability measurements identify deployment only through the directions they observe; we derive an exact ambiguity frontier when only small search widths are audited. Retaining candidate ranks and truth labels enables a constructive alternative: one audit can estimate reliability across all widths up to N. With known score percentiles, the minimax worst-coordinate mean squared error scales as (1 + log N)/T + N/M, capped at a constant, for expected budgets of T truth labels and M candidate observations. Matching lower bounds allow adaptive label acquisition, establishing that the distinct label and candidate costs are intrinsic to this experiment. An explicit design attains this order; a complementary record-based procedure supplies simultaneous guarantees without a known score distribution. Retrospective mathematical-reasoning and code-generation analyses show why search-dependent validation matters. In held-out CodeRM pools, a shared audit reduces the 95th-percentile maximum error across 100 widths by 58% and 40% relative to uniform labeling. These results turn structural ambiguity into a quantitative prescription for reusable validation.