Papers

Privacy-preserving learning is often motivated by the idea that protecting users' data can preserve trust and thus participation, improving utility in the long term. However, this claim has not been formalized so far. In parallel, performative learning provides a framework for studying learning systems whose deployment affects the data they later observe. In this work, we bring these two perspectives together and introduce performative privacy, where data leakage reduces future participation. We study a simple model where agents repeatedly contribute data for mean estimation but may leave the system when their data is leaked. Privacy is implemented through differentially private mechanisms, creating a trade-off between estimation noise and future participation. We show, through a theoretical study of the dynamics and numerical experiments, that a finite privacy budget can outperform non-private estimation in the long term when the feedback loop between leakage and participation is sufficiently strong. This provides first evidence that differential privacy can be optimal not only as a protection mechanism, but also from the perspective of long-term utility.

Circuit Representation Learning (CRL) offers a powerful paradigm to guide and optimize core Electronic Design Automation (EDA) tasks, but its practical adoption is hindered by the immense scale of industrial netlists and a failure to explicitly model register-level temporal dynamics. To overcome these barriers, we introduce DeepSeq3, a novel hierarchical framework that abstracts circuits into a two-level representation: fine-grained combinational subgraphs partitioned by flip-flops (FFs), and a high-level Super-Node Graph (SNG) that models the register-transfer structure. A dual Graph Neural Network (GNN) architecture learns representations at both levels, capturing local Boolean logic and global state transitions. Crucially, we introduce a state-centric pre-training scheme that predicts the reachability between FF states, endowing the model with a deep understanding of temporal behavior. Demonstrated on large-scale benchmarks, DeepSeq3's approach yields superior scalability and richer representations, reducing bounded model checking (BMC) solving time by 18% while guaranteeing correctness.

We study risk-averse decision making, in which an agent selects actions while being uncertain about the true system state. The risk is measured via optimized certainty equivalent (OCE) metrics, which generalize popular criteria such as mean-variance risk and conditional value-at-risk (CVaR). We characterize the optimal policy under known distributions, and show that it reduces to a prediction set-based solution for the CVaR. This provides an operational interpretation of conformal prediction-type prediction sets. For unknown distributions, we develop a data-driven calibration strategy, based on a synthetic model for the likelihood and held-out calibration data, yielding high-probability control of the OCE risk. The approach is evaluated on two wireless beamforming settings.

Biological knowledgebases like Reactome provide high-quality pathways that include biological elements' relationships and textual descriptions (metadata). The quality of such pathways is granted by manual curation, that presents, however, significant scalability challenges. Lately, numerous NLP tools have been proposed to cope with this issue, leveraging textual information to automatically expand biological knowledgebases. However, little exploration has been done so far to assess whether relationships among textual descriptions mirror higher order biological relationships. This study explores whether human-written descriptions in Reactome can be used to infer the experts' defined global hierarchical structure. To test this, we extracted from Reactome the Homo Sapiens hierarchy of pathways and their reactions (Reactome Hierarchy), and used textual metadata to reconstruct a Semantic Hierarchy, combining a sentence transformer model (SPECTER2) with a modified agglomerative nesting algorithm and a graph reconstruction algorithm. Quantitative (Laplacian Spectral Distance and Bootstrapping) and qualitative (global topological metrics) analyses confirm our hypothesis and indicate that the global hierarchical structure of pathways can be inferred by experts textual metadata.

Purpose - This study investigates the feasibility of restoring missing text caused by physical lacunae in damaged ancient manuscripts using language models. Methodology - The study proposes different scenarios to replicate real-world conditions. Language models of different architectures are applied according to their suitability to each scenario. We also propose several decoding strategies that further enhance performance and address the discrepancy between lacuna boundaries and the models' tokenization schemes. Findings - The results reveal that text restoration of these documents cannot be fully automated, but it can serve as a useful tool to assist paleographers in their work. Model performance varies greatly depending on which structural part of the document needs to be restored and whether the character length of missing text is available. Originality - This is the first study and to analyze model performance on formulaic and non-formulaic content and the impact of lacuna length awareness in manuscript restoration. Both are recurring challenges in paleographers' manual restoration work. Through systematic comparison and both qualitative and quantitative analysis of different models' performance under varying settings, this study offers a guideline for developing assistive tools to support paleographers.

AI agents are evolving into long-running computational entities that can invoke tools, maintain memory, and complete complex tasks across applications. In real-world settings, completing a task often requires humans and AI to take turns leading its execution. Such alternation depends on the seamless handoff of the work state of the task between humans and AI. Existing agent systems, however, provide humans and AI with separate work environments. AI agents must therefore rely on additional bridges to continue work: either developers build task-specific interfaces to access the work state, or users manually transfer relevant parts of it through screenshots or textual descriptions. Both approaches make handoffs costly and scale poorly. We propose Human-AI Co-inhabitation, a type of work environment that enables humans and AI to seamlessly take turns continuing work on the same task, and design and implement CrabOS to realize this concept. CrabOS represents the work state as natural-language-readable text objects shared by humans and AI, allowing both to access and manipulate it directly through the same auditable interface without bridges. Case studies show that CrabOS elevates support for complex tasks with alternating human and AI leadership from bridge-dependent application-level solutions to native operating-system capabilities, which provide a new foundation for developing and running AI agents.

FPGA-GPP heterogeneous systems combine software flexibility with the performance and energy efficiency of reconfigurable hardware. However, determining which application tasks should execute on the GPP or FPGA requires extensive expertise and design-space exploration, particularly when user objectives vary across latency, communication, resource utilisation, and power. This paper proposes Gen-TAS, a knowledge-grounded LLM framework for user-specific FPGA-GPP task allocation. By combining task-graph analysis with RAG, Gen-TAS grounds LLM reasoning in historical implementation knowledge and generates multiple explainable strategies tailored to the specified objectives. Human-in-the-loop selection and a deterministic backend connect LLM-generated decisions to reproducible FPGA SoC implementations. Experiments on CNN and SDR workloads across multiple LLMs demonstrate stable, requirement-driven allocation. Under latency-oriented objectives, implementations following the selected strategies achieve speedups of up to 2.45$\times$ and 92.53$\times$, respectively, relative to the corresponding all-GPP baselines while other objectives select strategies that trade some acceleration performance for FPGA-GPP communication, resource utilisation, or FPGA power.

Agitation fluctuates over short time horizons in people living with dementia, yet continuous physiological information for anticipating next-day risk is limited. We assessed whether contactless under-mattress signals from the preceding night inform next-day agitation risk and whether preserving minute-level temporal structure improves performance over conventional nightly summaries. We analyzed 423 patient-nights from 65 subjects in a specialized hospital dementia unit using two under-mattress sensing systems. A unified four-paradigm benchmark compared nightly handcrafted summaries, three-period handcrafted features, full-night sequence modeling, and sliding-window multiple-instance learning. Source-specific preprocessing and five-fold patient-grouped cross-validation were used, with performance estimated from pooled out-of-fold predictions. Evaluation included discrimination, calibration, fixed-threshold metrics, and a comparison of period-signal attribution patterns across two temporal models. Full-night sequence modeling achieved the highest discrimination (AUROC, 0.692; AUPRC, 0.849) and balanced accuracy (0.658). Both minute-level pipelines had higher AUROC than nightly summaries, but differences from three-period handcrafted features were uncertain. Cross-model attribution prioritized activity, heart rate, and respiratory rate during the core overnight period. Calibration remained limited. The preceding night's signals supported modest next-day risk discrimination, with minute-level temporal modeling outperforming nightly summaries. Prospective calibration and external validation are needed before use in individual care decisions. This patient-grouped benchmark identifies contactless overnight sensing as a promising biomedical engineering direction for agitation-risk research in hospitalized dementia cohorts.

A byte-level BPE tokenizer is an ordered list of merge rules, so applying only a prefix yields a vocabulary whose token identifiers are the first rows of the full vocabulary. This prefix nesting allows one language model to operate at several vocabulary sizes, use a control token to indicate the active size, and be deployed at any trained size by slicing its embedding and output head. We pre-registered five claims, including margins, seeds, contrasts, and a stop rule, and trained 30 models with 3.1M- and 10.6M-parameter bodies on 200M tokens each. Slicing is numerically exact: across 76 checks, a sliced model reproduces the restricted full model's logits bit for bit and removes 66% of deployed weights without changing latency. However, the shared model trails a fixed-cap specialist by 3.64% bits per byte at 32k against a 1% margin, and by 2.96% at 8k against a 2% margin. A 2x2 ablation separating the control token from output restriction finds that the token changes performance by +0.07% to +0.13%, with all intervals crossing zero, while output restriction costs +0.47% to +1.19%; the factors are substitutes rather than complements. Multi-cap training nevertheless improves robustness: under typographical noise, the same checkpoint degrades 12.5–15.4 points less in its fine mode and outperforms each fixed-cap specialist at that specialist's vocabulary size. A control with neither cap token nor output restriction is equally robust, attributing this benefit to multi-granularity training rather than conditioning. The per-cap penalty tracks each cap's share of training rows, yielding a falsifiable prediction for future work.

How much matrix rank is required to preserve every bounded value output of normalized softmax attention? We study the unrestricted maximum-row-\(\ell_1\) approximation rank \(r_\varepsilon(A)\), exactly the least rank achieving uniform error over all bounded vector-valued values. Row softmax exposes the intrinsic interaction \(C=P_m(\log A)P_N\), whereas invertible \(Q/K\) gauges leave \(A\) fixed while changing the Euclidean geometry of a chosen query/key factorization. We replace that coordinate-dependent description by a projective residual \(q(C-T)\) and an attained factor-radius size \(\kappa(T)\). For every rank-\(r\) retained interaction with \(\tau(T)<\varepsilon\), we prove $$ r_\varepsilon(A)\le \min\left\{ N,\; C_r\left( 1+\frac{\kappa(T)} {(\varepsilon-\tau(T))^2} \right)^{r/2} \right\}, $$ with the same unknown dimension constant as the underlying weighted Gibbs-row cover. The profile is gauge invariant, termwise no worse than native retained-subspace bounds at the same declared dimension, and has a worst-case sharp \(r/2\) size exponent at fixed \(r\) and \(\varepsilon\). We then measure \(r_\varepsilon(A)\) directly on learned attention using 9,978 certified brackets across BERT, GPT-2, Qwen2.5, and two ViT checkpoints; where certificates do not close, the optimum remains interval-valued. A pre-specified 2,302-cell held-out study further shows that the historical native-coordinate geometry block contains coarse, mostly head-level information but no detectable incremental information beyond a strong calibrated baseline. The new intrinsic descriptor is not evaluated in that study. Together, the theory and measurements distinguish an operator-intrinsic complexity control from a stronger empirical explanation that the learned-head evidence does not support.

Engineering agents that interact with external simulators may need to coordinate design modification with reacquisition of engineering evidence for the modified state. We ask whether first post-edit re-verification changes when explicit verification-cadence guidance is retained versus omitted while verification-relevant state/facts are held constant. Cadence-Guided (CG) retained an instruction to request a new simulation after a substantive modification, whereas Cadence-Omitted (CO) removed that instruction; neither condition used a hard gate. The study therefore measures instruction-conditioned post-edit verification-policy adherence rather than spontaneous recognition that prior evidence has become stale. Using DWSIM as the simulator backend and continuous valve-pressure adjustment, five Alibaba/Qwen models were evaluated on eight synthetic cases; each model-case-condition combination was executed three times via live API calls, yielding 120 evaluation slots per condition. Re-verification was observed in 94/120 CG slots versus 32/120 CO slots; cadence violations occurred in 26/120 versus 87/120; and bounded final success was reached in 95/120 versus 35/120. qwen3.5-35b-a3b showed minimal re-verification (1/24 in CG and 0/24 in CO) and no final success in either condition. Within this bounded protocol, explicit post-edit verification-cadence guidance was associated with more re-verification, fewer cadence violations, and more frequent bounded final success, supporting the treatment of verification cadence as an explicit interaction-protocol component.

Social power plays a fundamental role in shaping human interaction, yet computational studies of power remain limited to narrow linguistic and cultural settings. Existing datasets further lack the demographic and relational depth needed for robust cross-cultural analysis. To address this gap, we introduce a theoretically grounded framework for studying social power in naturalistic multilingual dialogue through movie screenplays. The framework integrates a schema informed by social science theory, a native speaker annotation pipeline refined through pilot studies, and a custom interface for scalable cross-lingual analysis. Using this framework, we constructed an initial corpus containing 15,836 annotated instances from 100 scenes in French and Egyptian Arabic movies. Our analysis reveals strong agreement on observable demographic and contextual attributes, while socially interpretive aspects, such as power asymmetry and intention alignment, remain more contested, highlighting the complexity of social power across cultures. We evaluated 6 Large Language Models (LLMs) and Multimodal LLMs on cross-cultural social power reasoning, finding persistent gaps between human and model agreement in relational and theory-of-mind reasoning. Our work introduces the first extensible multilingual framework for studying social power in dialogues and provides an initial evaluation setting for studying cross-cultural social reasoning.

We present CheXtriev, a graph-based, anatomy-aware framework for chest radiograph retrieval. Unlike prior methods focussed on global features, our method leverages graph transformers to extract informative features from specific anatomical regions. Furthermore, it captures spatial context and the interplay between anatomical location and findings. This contextualization, grounded in evidence-based anatomy, results in a richer anatomy-aware representation and leads to more accurate, effective and efficient retrieval, particularly for less prevalent findings. CheXtriv outperforms state-of-the-art global and local approaches by 18% to 26% in retrieval accuracy and 11% to 23% in ranking quality. The code is available at https://github.com/cvit-mip/chextriev.

Fine-grained credit assignment is a central challenge in reinforcement learning for long horizon LLM agents. Standard objectives often train from programmatically verifiable terminal rewards by broadcasting each sparse outcome to every action in a trajectory. Existing methods typically seek finer credit from the rollout side, constructing auxiliary trajectory signals or additional comparisons to estimate action importance. Although useful, these approaches still treat the verifier that judged success as a scalar reward, discarding its internal task structure. Our key insight is that many verifiable tasks already encode the relevant checks inside their terminal verifier. We propose VICT (VerifierInstrumented Credit Tracing), a training-time interface that exposes executable or evidence backed atoms and traces them back to actions through dependency-valid proof edges. VICT redistributes group-relative advantage only along those edges, shifting credit assignment from rollout-side inference to verifierside tracing. It preserves the original terminal reward, abstains when evidence is incomplete or ambiguous, and changes only the training-time advantage tensor, requiring no learned critic, process labels, branch rollouts, or inference-time verifier access. On ALFWorld and WebShop, VICT improves substantially over outcome-only training and achieves strong performance alongside recent fine-grained credit methods; ablations rule out dense atom rewards, final-commit credit, temporal proximity, and sparsity as sufficient explanations.

Real-time classification during language model inference is valuable for safety filtering, behavioral analysis, and model monitoring, but current approaches force a trade-off between accuracy and efficiency. Hidden-state probes are fast but limited: they are either not context-aware: operating on a single vector and cannot model interactions across positions; or they are very costly: having dedicated classifier models (Llama Guard, Qwen Guard, LLM-as-judge) or performing computation on hidden states for all tokens and then pooling the results (MultiMax). This shows an intrinsic trade-off between efficiency and accuracy. However, we find that the speculative-decoding module in recent LLMs can be repurposed for efficient high-quality classification. By appending a trained soft prompt at the end of the target sequence, we can repurpose the speculative-decoding module into a sequence classifier. At inference time in a speculative-decoding pipeline, the KV cache is already in GPU memory, so classification adds negligible overhead. We evaluate on four classification tasks across four models (Qwen3.5-4B, 9B, 27B, MiniCPM4.1-8B). Our small probes consistently outperform zero-shot GPT-5.4-mini and, on multilingual prompt safety, match or beat specialized 8B safety classifiers (Qwen3Guard-Gen-8B, Llama-Guard-3-8B) without running a full LLM.

Interpreting a CT scan means comparing structures on either side, judging how far apart organs sit, and knowing where each one belongs. Medical vision encoders are evaluated on diagnostic accuracy, or through assembled multimodal systems where a failure is hard to attribute, so it remains unclear whether their representations support any of this. We construct SPAR-Bench, eight probes over multi-organ abdominal CT that separate coordinate localization, relational reasoning, and spatial queries, and apply them to five architectural configurations and three medical foundation models, frozen and finetuned. Probes that ask for a comparison within the slice stay at chance, and neither pretraining scale, finetuning, nor architecture closes the gap. Probes that appear solved in domain fall to chance under zero-shot transfer, indicating that their accuracy reflects recall of canonical anatomy rather than computation over the image. Reading the same frozen features with a pooled head rather than the full set of tokens moves relational recovery from 0.7% to 67.8%, so pooled probing understates what a representation holds. Questions the encoders answer well are answered at chance by four open-weight MLLMs. Our results suggest these encoders carry a map of where organs usually lie, and little of the machinery for comparing structures within a particular patient. Code and data will be available at https://spar-bench.github.io.

Recent progress has been made in 3D Gaussian representation for reconstruction, generation, and physical simulation. However, current approaches mainly concentrate on physics-based dynamic generation of solid objects and only handle single-phase collision interactions. We introduce VersaGauss, a unified framework for generation, simulation, and rendering that supports versatile physics-based dynamic generation, particularly for multiphase interactions. Our system takes a few images as input and produces a realistic, physics-driven 3D dynamic scene with multiple objects. To optimize the Gaussian kernel distribution, we develop a particle pruning algorithm. We also propose the Coupled Multiphase Point Method (CMPM) to effectively model and generate multiphase interactions. Additionally, harmonic interpolation within CMPM and a Gaussian evolution strategy are introduced to achieve realistic fluid rendering. Extensive experiments demonstrate that our framework can simulate interactions among various materials such as fluid, rubber, sand, snow, and others. Code is available at https://github.com/Elowen-surj/VersaGauss.

Existing API-only prompt optimisers are often described as interpretable, but in practice, this usually means only post-hoc inspectability: each iteration still rewrites the prompt as one opaque string, leaving a trace of full-prompt diffs rather than localisable, machine-readable edits. This paper introduces SEPO (Structural, Evidence-grounded Prompt Optimization), a multi-trajectory prompt optimiser centred on edit-effect lineage feedback. Rather than treating each iteration as an isolated whole-prompt rewrite, SEPO locally edits stable, typed units in a two-layer prompt schema, links the target and realised structural operations of each edit to the examples it newly fixes or breaks, and carries this edit-effect record forward to guide later architect calls on the same search branch. This makes prompt optimisation addressable, attributable, and actionable. Across a 14-task held-out suite, SEPO improves over the strongest baseline, GEPA, by 3.1 pp on Llama-3.1-8B-Instruct and 2.2 pp on Qwen3-8B, reaching 61.9% and 73.3% macro accuracy. SEPO also lies on both the optimisation-time and test-time Pareto frontiers, spending 2.9M optimisation tokens versus 4.1M for GEPA and producing prompts over 5x shorter.

Complete your ad view and grab a 5-cent bonus! In incentivized advertising, a platform promises users a bonus before observing downstream ad revenue, encouraging them to click and complete ads. It must balance the incentive promised in advance against the revenue realized afterward: insufficient incentives forfeit monetization opportunities, whereas excessive incentives reduce net profit. Because current incentives may also shape user expectations and future engagement, incentive allocation is a sequential decision problem with delayed revenue, cost sensitivity, and carryover effects. Existing work has not studied decision-making algorithms for this setting. Auto-bidding assumes available ad opportunities, while targeted promotion optimizes incentives outside the ad monetization pipeline. We formulate the problem as an MDP and develop an offline model-based RL framework for cost-controllable sequential incentive allocation. It learns a world model of user feedback and ad revenue, then performs conservative policy optimization. An independent counterfactual scorer evaluates each learned policy on held-out logs, enabling pre-launch selection without costly online exposure. Experiments on large-scale industrial data and online A/B tests show that the scorer provides a stable offline signal. The deployment path from causal inference to offline RL and then Offline-MBRL further validates the framework: MB-IQL improves per-user net profit by 7.96% over TD3+BC, whereas reverting to plain IQL reduces it by 6.56% (both \(p<0.0001\)).

Multimodal search agents extend parametric knowledge with newly emerging and long-tail evidence from the open web. Yet many existing agentic search environments often expose retrieved evidence only as text and omit tool-returned images from subsequent context, reducing visually grounded trajectories to text-only reasoning. Long-horizon interaction also compounds tool-call, response-length, timeout, and budget failures, which can discard salvageable trajectories, waste rollout computation, and disturb policy updates. To address these issues, we introduce WeAgent-Harness, a multimodal agentic harness that supports native text-vision interaction and runtime recovery. Retrieved images receive persistent disk references, allowing the model to inspect, process, and cite them throughout the trajectory. Based on this harness, we develop WeAgent-MMSearch, an integrated system spanning data construction, agentic post-training, and multimodal rollout. For data construction, a strong MLLM uses WeAgent-Harness to discover, synthesize, and verify MMSearch-style tasks and collect expert trajectories. During post-training, our Failure-Aware GSPO (FA-GSPO) recovers salvageable abnormal rollouts and filters invalid ones to improve bounded multimodal planning and search. We also introduce VisTarget-Bench, a 150-task human-verified benchmark that pairs each question with a held-out target image, distinguishing image-retrieval failures from visual-perception failures. Evaluation on VisTarget-Bench and seven public benchmarks shows that agentic post-training improves the average score by 19.22 points, enabling our model to outperform similarly sized open-source models and rival models with roughly ten times its parameter count.

Large Vision-Language Models (LVLMs) remain prone to hallucinations, producing responses that are irrelevant or inconsistent with the multimodal input. Existing mitigation methods mainly rely on external supervision, output calibration, or attention regulation, leaving the internal representation dynamics of autoregressive generation underexplored. We identify an inference-time failure mode in which cross-modal representations degrade across decoder layers and drift across generation steps, destabilizing token prediction and increasing hallucination risk. We propose Dynamic Alignment Compensation (DAC), a training-free inference-time method that detects representation divergence and selectively applies lightweight residual compensation. DAC combines Layer-wise Semantic Compensation to mitigate inter-layer degradation with Sequential Semantic Correction to constrain temporal drift. Experiments on nine hallucination-focused and general-purpose multimodal benchmarks across multiple LVLM backbones show that DAC consistently reduces hallucinations while maintaining strong overall performance.

Artificial intelligence has strong potential to support clinical decision-making, yet its adoption in healthcare remains limited due to a lack of trust. Uncertainty estimation can signal unreliable predictions, and explainable AI (XAI) can clarify how predictions are made but existing methods treat them separately, providing no feature-level insight into why a prediction is uncertain or which tests to prioritize to reduce it. To address this gap, we propose explainable uncertainty estimation, which unifies uncertainty estimation and XAI to both quantify uncertainty and explain feature-level contributions. We introduce the Expected Gradients Reconstruction Uncertainty Estimate (egRUE), which incorporates prediction explanations into its uncertainty computation and decomposes uncertainty into feature-wise contributions. We prove theoretical properties of egRUE and show through experiments that it improves reliability and interpretability compared to existing methods. A user study with medical experts further demonstrates that egRUE's explanations improve calibrated trust over uncertainty scores alone, increasing confidence in correct predictions and reducing confidence in incorrect ones. By combining prediction uncertainty with feature-level explanations, egRUE strengthens decision-making support in safety-critical healthcare settings, clarifying both when predictions may be unreliable and which features drive that uncertainty.

Text simplification aims to make complex texts easier to understand while preserving their original meaning. Recent large language models can perform simplification through prompting, but it remains unclear whether adding explicit linguistic information about sentence complexity to the prompt improves their outputs. We investigate this question for multilingual sentence-level Easy-to-Read simplification in Catalan, Spanish, and Italian. Using Qwen3-8B, we compare a baseline prompt, a gold-cue prompt enriched with gold linguistic cues, and a predicted-cue prompt enriched with automatically predicted cues. We evaluate the outputs using SARI, BLEU, chrF, and BERTScore, and complement this evaluation with a manual qualitative analysis. Predicted-cue prompting obtains the best overall scores across all four metrics, although the gains over the baseline are small. Gold-cue prompting does not consistently improve over the baseline, and results vary across languages. These findings indicate that cue-based prompting can influence multilingual Easy-to-Read simplification, but its benefits are modest, metric-dependent, and language-dependent.

LLM agents have become a new class of software user, but every surface they work through was designed for someone else. Pages are built for human eyes, which can skim and ignore; tool schemas for programs, which pay nothing to carry definitions they never call. An agent has neither luxury: it re-reads, and pays again for, everything it is shown on every turn. We present String, an open-source runtime that gives this user an interface of its own and treats the job as an operating-systems problem. Tool knowledge moves out of the agent's context and into a common layer that renders it back one view at a time as Markdown. A single SFMD (String-Flavored Markdown) document declares an application's views, typed actions, navigation, and credentials, and the runtime handles discovery, validation, execution, state, and secrets behind two core verbs: /open to see and /act to do. Web and app turn out to be two renderings of one architecture: an SFMD site serves styled HTML to browsers and the raw document to agents, so one grammar reaches apps, files, shells, and the web, even legacy HTML, with no per-site integration. Views stay partial by design, and the staging is causal: disclosing one tier of detail a single turn too early costs up to 23 accuracy points, while proper staging drops wrong-action selection from 28% to 2%. Privilege follows provenance: a remote page may call HTTP but never the shell, and caller-supplied text never expands a stored secret. On an 87-task benchmark that pairs each task with curated skills, operationalizing those procedures as on-demand String apps yields comparable aggregate success across six models from frontier to small (+1.3pp) while using 33.5% fewer tokens among completed episodes, and the resident interface stays a constant 53 tokens at any catalog size. We report the design, the evaluation, and what three months of production use taught us.

Large language models increasingly author Infrastructure-as-Code (IaC), where one insecure default is provisioned straight into production. Prior evaluations report vulnerability counts for models only, and so cannot say whether models are worse than the engineers they assist. We present GenIaC-SecBench: 100 deployment scenarios across 12 model configurations from six vendors, open and closed weights, yielding 1,196 artifacts scanned by three policy engines (Checkov, Trivy, KICS) at complete coverage. Crucially we scan 634 human-authored IaC templates with the identical toolchain, giving the first size-matched human security baseline for this task. Vulnerability density is strongly inverse to artifact size (Spearman $\rho=-0.55$, $p<10^{-77}$), so unmatched comparisons measure size, not security. Size-matched, every configuration exceeds the human baseline at $3.21\times$ to $3.87\times$, and the gap widens as tasks get simpler ($4.9\times$ at one resource, $1.4\times$ at twenty or more). A majority of scenarios prescribe a security state rather than specifying function alone, so we stratify by prompt class: pooled the gap is $3.50\times$, and excluding every scenario that explicitly requests an insecure configuration still leaves all configurations above baseline ($2.4\times$ to $4.2\times$). The corpus cannot isolate unprompted default posture, and we say so. Decomposing "reasoning" into standard generation, prompted chain-of-thought, and vendor extended-thinking APIs, extended thinking beats prompted CoT ($-12.0\%$, $p=0.0013$) while prompted CoT alone is indistinguishable from standard ($-1.3\%$, n.s.); it consumes under $1\%$ of the output budget, bounding the effect. Two negative results: more deployable models are not more vulnerable ($r=0.158$, $p=0.625$), and complete-case Friedman is uncomputable here, motivating Skillings-Mack. All code and data are released.

Knowledge-intensive reasoning requires Large Language Models (LLMs) to ground answers in provided evidence. When evidence is insufficient, it is desirable that models abstain rather than confidently generating unsupported answers. Existing abstention methods rely on uncertainty estimation or evidence sufficiency checks, but neither tests whether the reasoning process for generation, driven by the interaction of provided evidence and the model's internal memory parameters, is actually grounded in the evidence. A key contributing factor is that entity mentions in context activate memorised associations, causing models to generate plausible responses ungrounded in evidence. We propose Twin Worlds (TW), a framework for improving reliability in knowledge-intensive reasoning through equivariance-based abstention: unlike invariance, which requires outputs to remain unchanged, equivariance requires outputs to transform correspondingly under entity substitutions. A model grounded in the evidence should produce answers that shift consistently when entities are substituted while their relations are preserved. TW constructs multiple worlds via typed substitutions of the original input that preserve relational structure while reducing parametric priors, and uses equivariance violations as an abstention signal. Across four benchmarks and three model backbones, TW identifies when answers are not reliably grounded in the provided evidence and outperforms uncertainty- and sufficiency-based baselines.

Trajectory-level credit assignment can localize which module of a tool-using LLM agent causes failures using only verifiable signals. We ask whether such failure credit should route a fixed zeroth-order/evolution-strategies (ZO/ES) perturbation budget. Across a synthetic environment and frozen Qwen2.5-1.5B/3B and SmolLM2-1.7B agents, three task families, six allocation schemes, a credit-noise sweep, paired seeds, and exact sign-flip tests, we find no statistically detectable improvement over uniform allocation in any on-pool comparison (no gain of at least 2 percentage points). The joint soft-plus-sigma scheme is equivalent to uniform within a +/- 0.02 AUC margin on 1.5B and 3B; concentrating the full budget on the credit argmax is marginally equivalent on 1.5B, where that module is the verified bottleneck, and significantly worse on 3B. Inverse-propensity debiasing does not rescue routing, and misrouting costs up to -0.074 AUC in-house and -0.118 end-to-end on the BFCL-derived family. Across six fixed-step schedules, loss is linear in bottleneck starvation rate (R^2 = 0.94, descriptive), and a preregistered credit-free coverage floor removes detected harm. Matched-budget burst and step-compensating catch-up schedules are consistent with harm arising from insufficient cumulative parameter movement rather than update frequency. Our primary estimand is optimization efficiency on a fixed task pool. On unseen BFCL functions, the study's one exception is that soft routing exceeds uniform on held-out endpoints (+0.047, p = 0.031, n = 6). A plausible but untested reading is that routing-favored caller improvements transfer while uniform's on-pool gains reflect a synthesizer behavior specific to our harness. We report this exception explicitly and document three failure modes that can silently invalidate ZO/ES experiments on frozen LLMs.

Flow matching enables likelihood-free training, yet alignment methods increasingly reuse conditional flow matching (CFM) losses as endpoint negative log-likelihoods (NLLs) and their old/new differences as log-likelihood ratios. We characterize when these substitutions are valid. For linear Gaussian paths, we exactly decompose endpoint NLL into entropy, a weighted CFM objective, an interior velocity–score residual, and a boundary residual. Thus CFM-only estimates and differences are exact only when the corresponding residuals cancel. At the off-policy population optimum, ordinary CFM is not generally a pointwise NLL estimator, whereas \(w_{\mathrm{sc}}(t)=(1-t)/t\) removes the interior residual; this positive result does not extend generally to training or on-policy alignment. On-policy log-ratios can remain biased even for identical endpoint laws or after surrogate optimization. Experiments across dimensions, distributions, and geometries support these conclusions and the mechanisms that make inexact ratios useful. **More broadly, the decomposition provides a theoretical basis for adapting likelihood-based LLM methods to flow matching, while distinguishing exact substitutions from controlled surrogates.**

This paper proposes a layer bit allocation method for Gemma-3-1B, formulating the problem as performance maximization (latency decrease) given a degradation budget constraint (allowable level of generation quality loss). This approach is different from time- and resource-consuming uniform layer quantization methods that are used in the literature (like GPTQ or AWQ) or allocation methods without proven performance-accelerating effect (like MixLLM or TorchAO). The layer sensitivity profile resulting from our prior work SA-PTQ is applied using the activation pass-through mode inside TensorRT-LLM. For each layer precision is determined individually in blocks, according to a grouping introduced in the prior step (5+5, 10+10, all26), differentiating the contribution of FFN, Attention, and lm_head to the overall speedup. The clock speed was measured for 13 W8A8 variants on an RTX 5090. We find that for FFN and lm_head the time cost of quantization/dequantization is compensated for by the use of integer arithmetic, while for short context lengths, the opposite holds true for Attention: an additional step of quantization slows execution down. We propose a manual implementation of SmoothQuant for TensorRT-LLM which was necessary due to export failures, unavailable for lm_head. The best solution found under joint consideration of all three criteria with minimal degradation was FFN 5+5 with lm_head, providing an 11.0% reduction in latency with negligible quality loss (98.90% Top-1 agreement, +0.85% perplexity degradation). With acceptable quality loss for FFN all26 + lm_head, a speedup up to 19.1% was found possible. We suggest further optimizations: fused attention kernels in INT8, KV-cache quantization, using FP8 instead of INT8 and partial Attention quantization analogous to FFN.

Existing conversational plant-phenotyping platforms are difficult for plant scientists to use and lack the reliability scientific research demands: failed analyses are reported as valid measurements rather than flagged as missing, statistical tests run without checking assumptions, predictions carry no uncertainty estimate, and specialised hardware limits accessibility. We present PhenoIntel, a lifecycle-aligned multi-agent web platform that turns the full machine-learning workflow into a reliable, user-friendly phenotyping system. Nine specialised agents divide the analysis into stages, from image collection through model selection, inference, and reporting, rather than handing the whole task to one AI manager. Independent checks separate these stages, and every agent reads from and writes to one shared, fixed-structure record, so an inconsistent output from one stage is caught before it reaches the next. Uncertainty is matched to each model family, conformal prediction, detection-confidence spread, or Monte Carlo Dropout, rather than applied uniformly, and quality thresholds adapt to crop and task instead of one global cutoff. When no suitable model exists, PhenoIntel can propose, validate, and integrate a new one on its own. The model repository spans ten trained models across five crops and four imaging modalities. Classification models reach Macro F1 of 0.78-0.996; object-detection models reach 0.96 mAP@50 with a 54% reduction in counting error over an unoptimised baseline; and a temporal model reaches held-out Macro F1 of 0.7050. PhenoIntel runs in a browser on standard hardware, requiring no GPU, and a 1,200-test automated suite confirms complete pipeline execution. Every result carries calibrated uncertainty, validated statistics, and FAIR-compliant provenance, a combination existing conversational phenotyping tools do not offer.