Papers

We measure the rate at which code RL environments accept incorrect solutions as correct. On a 49-task sample of SWE-bench Verified, 28.5% of tasks have test suites weak enough that a Docker-verified incorrect patch passes them. On 20 R2E-Gym tasks across 6 repositories, the same pipeline at single-shot exploit generation yields 25.0%. A random-effects meta-analysis over 134 frontier model submissions to SWE-bench Verified finds, within the same human-rated difficulty stratum, model Pass@1 is +14.14 percentage points higher on flagged-hackable tasks than on robust ones (95% CI [+11.80, +16.48]; one-sided p < 10^-6; I^2 = 0%; 123 of 134 models positive). We then describe a procedure for hardening the broken tasks. An inline LLM judge with a Docker gold-sanity gate runs each generated test against the gold solution before the judge is consulted. On the 11 broken tasks in the audit, the gate flags 65 of 105 decisive LLM-generated tests as failing on the gold patch itself, a 61.9% per-augmentation defect rate the LLM judge alone misses. With diversity-biased retry, the loop converges 9 of 11 tasks to a gated upgrade.

For thirty years, quantitative finance has paid a costly two-language tax: models researched in Python are rewritten in C++ for production, often introducing numerical discrepancies. GPU-accelerated deep learning exacerbates this problem, as nondeterministic floating-point reductions can produce drift in long backtests, challenging regulatory reproducibility and auditability expectations. This article surveys Mojo, Modular's 2026 Python-like systems language, as a structural response for capital markets engineering. While closing the Python-to-C++ performance gap, Mojo uniquely combines native interoperability with the low-level systems control required to construct bit-exact deterministic kernels. Its MLIR compilation infrastructure further allows a single codebase to target scalar, SIMD, multicore, and GPU execution, reducing the translation bottleneck between research and production. We benchmark four core financial AI workloads: Monte Carlo option pricing, LLM sentiment inference, multi-asset backtesting, and portfolio Value at Risk. On Apple Silicon, Mojo demonstrates 20x to 180x speedups over pure Python on directly measured kernels; larger-scale GPU workload results are projections calibrated from published benchmarks. Alongside transparent performance data, we introduce mojo-deterministic, an open-source library of reproducible reduction kernels, and provide a candid assessment of the problems Mojo does and does not yet solve.

Research on artificial intelligence and democracy has grown quickly over the last decade. A shared conclusion in this literature is that AI does not create new democratic problems so much as it makes old ones worse. We now see this across information ecosystems, in elections, and in public administration. However, despite growing evidence, we lack a clear way to prioritize risks in this area, compare them across domains, and identify where democratic control is most likely to break down. So, our problem is: How can we systematize the problems that AI systems pose to democratic processes? This paper argues that principal agent theory may fit the task. In many phases of democratic systems, principals delegate key functions to AI systems and their providers without really being able to monitor how these systems operate or the outputs they produce. Treating AI as a delegation problem helps identify accountability gaps and other governance failures. Most importantly, as we shall illustrate, it provides metrics for empirical assessments of AI impact on democracy. As a second analytical element, we draw on the NIST AI Risk Management Framework and its seven characteristics of trustworthy AI, which supply substantive criteria for evaluating delegated tasks. Operationalized across the three domains through measurable indicators and domain specific trustworthiness criteria, we propose an analytical framework that centers on institutional assessability as the central condition for democratic control over AI. However, we stress that how severe a harm is, and how much risk is acceptable, are evaluative judgments that current methodologies neither acknowledge nor operationalize. This becomes acute when such evaluative judgments are (silently) delegated to private vendors. We identify this as a strong limitation left for future work.

Robust deep learning under heavy-tailed and impulsive noise remains challenging because conventional losses such as mean squared error (MSE) exhibit unbounded sensitivity to outliers. Although correntropy-based objectives improve robustness, many conventional formulations rely on fixed kernel parameters that must be empirically tuned and remain static during training. To address these limitations, we propose an Adaptive Log-Correntropy Loss (ALCL), a heavy-tailed loss formulation that adaptively learns its robustness geometry during optimization. ALCL introduces a logarithmic residual model whose shape and scale parameters are learned jointly with network weights through differentiable reparameterization. This yields a heavy-tailed robust optimization objective whose psi-function is formally bounded and redescending, allowing the loss geometry to adapt dynamically to evolving residual statistics while suppressing extreme outliers. Comparative experiments on four widely used benchmark datasets spanning grayscale and red-green-blue (RGB) image data under mixed heavy-tailed and impulsive noise demonstrate that ALCL consistently outperforms MSE and optimally tuned generalized correntropy losses in both reconstruction fidelity and downstream classification accuracy. While performance differences remain small under low-noise conditions, under high-noise regimes ALCL improves median accuracy by up to 4.75% on grayscale benchmarks and 4.51% on RGB datasets, with reduced variance across runs. These results demonstrate that adaptive robustness through joint learning of loss parameters provides a computationally efficient alternative to static correntropy-based losses for deep learning in non-Gaussian environments.

This work explores the use of artificial intelligence in mobile robotics to achieve autonomous detection and pose estimation of load carriers for automated pickup. A deep neural network is designed to recognize predefined landmarks on the carrier from RGBD data; these landmarks are then used to compute the carrier's pose. The network operates directly on RGBD images to estimate landmark positions, which form the basis for determining the carrier's location. The approach is validated in extensive experiments and comprises both software and hardware implementations. A deep learning-based framework is presented to detect load carriers and estimate their pose for use with autonomous logistics vehicles. Our method uses a convolutional neural network to identify characteristic reference points on the carrier from RGBD input and computes its pose by combining these inferred landmarks with prior geometric knowledge. Experiments show that the resulting accuracy is sufficient for reliable load carrier detection in industrial environments, confirming the suitability of the method for autonomous intralogistics applications.

The path toward autonomous software engineering is currently bottlenecked by a severe deficit of diverse, large-scale trajectory data. We address this by introducing \ourdataset, an expansive dataset of 207,489 agentic trajectories spanning nine programming languages (Python, Go, TS, JS, Rust, Java, PHP, C, C++). Sourced from 20,000 real-world PRs via OpenHands and SWE-agent harnesses, the dataset utilizes a hybrid-reasoning synthesis: Minimax-M2.5 generates trajectories with explicit "thinking" processes, while Qwen3.5-122B provides high-quality "non-thinking" traces. Filtered for permissive licenses (MIT, Apache, BSD) from SWE-rebench-V2, this data facilitates the training of models capable of long-horizon reasoning. We validate the dataset by fine-tuning the Qwen3-30B-A3B series (Thinking, Instruct, and Coder). The best performing model achieves resolve rates of 61.7% on SWE-bench Verified, 57.1% on SWE-bench Multilingual, and 36.8% on SWE-bench Pro. These results establish Open-SWE-Traces as a premier resource for distilling human-level software engineering capabilities into efficient, open-source agentic LLMs.

Many games rely on storytelling combined with systems that track levelling, NPC behaviour, and consequence simulation; bridging tightly-authored narrative with deeply-simulated worlds – most acute in sandbox and open-world settings – has been prohibitively expensive. LLM-driven worlds open a new path: a single harness can coordinate numerical state, narrative voice, storytelling pacing, and rule logic together. Realising this requires the LLM system to sustain a persistent world (who is where, what has just happened, what is currently true), which today's deployed systems do not: the narrative voice asserts state in free prose without any validated representation, so a fully autonomous game engine remains infeasible. We treat this as an architectural choice, not a limitation of language models, and report work in progress on a framework – orchestrated reality – that makes the world a canonical object owned by a singleton orchestration agent analogous to the tabletop-RPG Game Master (GM). We formalise an LLM-driven game world for a human player as a Parameterized-Action POMDP: state is a tree of canonical JSON entities, actions decompose as $a=(k, x_k)$ (a discrete intent kind plus structured JSON parameters), the agent observes only a narrative projection $o=O(s)$ of state, and the transition kernel $F$ is an LLM-driven Plan-Diff-Validate-Apply (PDVA) pipeline that commits schema-validated, content-hashed JSON deltas. We give the formal model, a JSON-state example, a worked single-turn example, and a catalogue of 15 illustrative incidents drawn from a real deployment showing the framework in action. Empirical validation through a planned human player study – together with multi-NPC concurrent agency and deployment as an RL environment – is situated as future work.

Large language models have achieved impressive performance on reasoning tasks spanning mathematics, science, programming, and commonsense inference. Despite these advances, their reasoning processes remain largely latent, making them difficult to interpret, verify, replay, debug, and transfer across domains. Existing approaches such as chain-of-thought, tree-of-thoughts, graph-of-thoughts, and tool-augmented reasoning expose intermediate reasoning artifacts but typically lack explicit execution semantics, formal state representations, and verifiable reasoning structures. We introduce Theorem-Grounded Execution Ontologies (TGEO), a framework that models reasoning as an executable state-transition process rather than a sequence of generated tokens. Given an input problem, TGEO identifies relevant theorem families, binds the problem to a domain ontology, discovers semantic objects, instantiates states and operators, constructs predicates and contracts, and synthesizes an executable reasoning graph. The resulting graph provides an interpretable, replayable, and auditable representation of reasoning in which every state transition, operator application, and validation step is explicitly represented. TGEO integrates five architectural components: (1) theorem-grounded reasoning priors, (2) executable ontologies, (3) operator-mediated state transitions, (4) predicate and contract-based execution validation, and (5) architectural auditing and failure localization. We evaluate TGEO on theorem-intensive reasoning tasks derived from mathematical benchmark domains and a curated Golden Execution Suite. Our findings demonstrate the value of executable reasoning representations for interpretable, verifiable, and reproducible AI reasoning systems.

This work investigates the ability of large language models (LLMs) to generate mathematical equations from scientific texts. Prior work faces challenges in unstructured grounding, multi-equation dependency, and humanaligned evaluation. To this end, we construct a dataset of AI research papers, pairing contextual passages with ground-truth equations and variable descriptions. We develop an explainable equation generation workflow and evaluate it across diverse open- and closed-source LLM backbones. We introduce an evaluation protocol combining automatic metrics, LLM-based rubrics, and human judgments to assess accuracy, explainability, and human-LLM alignment. Results indicate that LLMs perform moderately on lexical- and syntactic-based similarity, while struggling with semantic accuracy. Comparisons between LLM-based evaluations and human judgments reveal limited alignment, highlighting challenges in using LLMs to assess equation quality. These findings offer insights for improving equation generation models and developing more reliable evaluation methods for scientific text. We provide code and data for reproducibility.

Entity-aware document retrieval uses query-associated entities as ranking signals, assuming that semantically relevant entities are also useful retrieval signals. We show this assumption is insufficient- and explain why. Unlike terms, which are ground-truth observations, entity links are hypotheses produced by an imperfect linker: an entity can be topically central yet provide no discriminative signal if the linker fires indiscriminately across relevant and non-relevant documents. We formalize this as a distinction between Conceptual Entity Relevance (CER)- whether an entity is topically related to a query- and Observable Entity Relevance (OER)- whether its observed presence in a collection discriminates relevant from non-relevant documents. Across four collections and annotation sources including human entity judgments, CER and OER exhibit near-chance agreement ($\kappa \approx 0$), while OER operationalizations agree substantially ($\kappa \approx 0.5$), confirming CER as the systematic outlier. CER-based supervision selects topically plausible but weakly discriminative entities, pruning fewer than 4% of non-relevant documents on some collections. Aligning supervision with OER improves non-relevant pruning by up to 10x and open-world MAP by 0.051 over BM25. Our findings motivate a shift from conceptual to observable notions of entity relevance in entity-aware retrieval.

Translating deep learning models from PyTorch's flexible, object-oriented design to JAX's functional, stateless setup is usually a manual and error-prone task. Automated migration is challenging because Large Language Models (LLMs) struggle with strict and dynamic API alignment and are prone to mistakes for exacting operations. We propose a fully autonomous system that combines In-Context Learning (ICL) with oracle-driven self-debugging. First, we curated an ICL context that serves as a strict reference for idiomatic JAX styling and test case generation. Second, instead of depending on the LLM to deduce mathematical outputs, we run the source PyTorch modules to get their actual dynamic tensor states. This creates an unchangeable execution oracle. We then use an autonomous agentic loop to synthesize tests based on the oracle data. The test cases are executed repeatedly, and the traceback is sent back to the LLM for self-correction. Ablations show that combining ICL references with oracle grounding and self-debugging greatly outperforms pure instructional and basic agentic baselines. This improvement does not add an excessive computational overhead. Our lightweight pipeline achieves 91% numerical equivalence (compared to baseline: 9%, instruction + self-debugging: 27%) on neural modules, providing a highly reliable, scalable blueprint for cross-framework migration. This has been validated across several state-of-the-art models including SAM (segment anything), T5, Code Whisper amongst others showing high numerical equivalency. Code: https://github.com/AI-Hypercomputer/accelerator-agents/tree/main/MaxCode

Aligning neural activity across subjects offers the promise of discovering shared computational principles and generalizable decoders. However, traditional alignment methods require shared stimuli across subjects, a constraint that limits applicability to naturalistic paradigms with limited or non-overlapping data. We introduce a Multi-Encoder-Decoder Variational Autoencoder (MED-VAE) that achieves cross-subject alignment without shared stimuli by anchoring representations to a common scaffold provided by a pretrained ANN. Using the Natural Scenes Dataset, we show that MED-VAE creates common latent spaces with superior semantic organisation, achieving higher cross-subject alignment than common methods while maintaining robust generalisation to held-out stimuli where traditional methods degrade. Reconstructing from these common spaces back to each subject's original neural space, MED-VAE preserves equal stimulus-driven signal in its cross-subject latent space. Finally, we show that this superior alignment directly enables cross-subject neural prediction, as demonstrated via cross-subject image decoding. In summary, we introduce a framework to identify generalisable common subspaces for cross-subject predictions and downstream tasks, demonstrated here for visual cortex responses to static images.

Activation monitors – lightweight probes trained on a language model's internal representations – are an increasingly common layer in deployment safety stacks. Deployed models however are rarely static: they are quantized, fine-tuned, adapted with LoRA, or served with merged adapters while the monitor remains frozen. We present the first systematic test of whether this implicit contract holds: whether activation monitors trained on a base model remain reliable after these routine model updates. Across multiple safety-relevant monitors, model depths, update families, and open-weight models, we find a sharp split: quantization-style updates largely preserve frozen probe performance, while fine-tuning-style updates frequently make probes stale. Fragility is highly monitor-dependent, with privacy/PII probes most affected and refusal-compliance probes comparatively stable, showing that retraining a behavior need not stale its corresponding monitor. QLoRA is especially damaging despite NF4 quantization alone being relatively benign, suggesting that quantization becomes riskier when combined with adaptation. We further show that degradation is predictable from pre-deployment features, enabling revalidation budgets to be triaged toward the monitors most likely to fail. Finally, we test repair strategies and find that cheap label-free activation realignment repairs every repair-relevant stale cell, with none requiring score calibration, few-label heads, or labeled retraining. These results suggest that fine-tuning should trigger activation-monitor revalidation by default, with prediction triaging which monitors to check first and label-free realignment as the default repair.

With large language models (LLMs) increasingly applied to mathematical reasoning, formal proof assistants such as Lean can be leveraged to verify reasoning outputs with machine-checkable rigor, enabling use cases such as answer selection in test-time scaling with K sampled candidate answers. However, employing Lean requires that LLM outputs, originally in natural language, first be formalized. Existing Lean-based answer-selection work uses an autoformalization model to generate a formal statement in Lean for each candidate answer independently, incurring a significant computational cost. We propose BASE, a base-and-edit pipeline that formalizes a single base candidate per problem and derives the remaining K-1 statements by editing the answer expression in place. To facilitate this, we train a rewriter model LEANSCRIBE to localize the answer in the base formalization and generate a reusable edit function for the other K-1 candidates. BASE simultaneously improves selection accuracy and reduces formalization cost - a Pareto improvement that holds on all 12 (dataset, solver) configurations across four benchmarks and three solvers, cutting autoformalizer calls by about 5x at K=8, with the reduction expected to become larger as K grows. Code is available at https://github.com/ucr-rai/base-and-edit.

Federated fine-tuning of large language models using parameter-efficient methods such as LoRA enables privacy-preserving adaptation of foundation models. Heterogeneous hardware resources introduce challenges, as clients with different adapter ranks cannot be directly aggregated. While existing methods enable aggregation under heterogeneous ranks, they fail to control how information is distributed across rank dimensions, leading to suboptimal use of shared low-rank representations. Instead, we propose PreLort: a nested low-rank formulation for federated LoRA that organizes adapter dimensions into a prefix hierarchy. Our approach ensures that lower-rank dimensions encode task-relevant information, while higher-rank dimensions capture additional capacity. Building on this, we introduce (i) a segment-wise aggregation rule that averages only over clients contributing to each rank segment, avoiding dilution from zero-padded lower-rank clients, and (ii) a prefix-nested training strategy that optimizes each adapter under multiple rank truncations, encouraging useful signal to concentrate in low-rank prefix dimensions. Together, these components encourage a consistent low-rank prefix capturing the most task-relevant information, while higher-rank dimensions learn additional capacity. This allows low-rank clients to benefit from richer information contributed by higher-rank clients, as prefix dimensions are consistently learned and aggregated. Experiments demonstrate that our method consistently outperforms prior heterogeneous federated LoRA methods in accuracy and ROUGE-L, while achieving lower or comparable perplexity across multiple base models.

Neural networks are used as generative surrogate models for scientific discovery, which are trainable approximations of scientific simulations. These models enable users to replace time-consuming numerical simulations with learned alternatives, providing quick solutions. However, high-fidelity generative surrogate models require massive training datasets, which can create storage and I/O challenges. Lossy compression is a promising way to reduce this burden, but compression errors may affect the model quality in subtle ways, making it challenging to quantify their impact. In this work, we examine how lossy compression of training data impacts the quality of generative surrogate models. We begin by characterizing the uncertainty inherent in training neural networks, showing that identical training configurations can produce different models. By exploiting this variability, we propose a method to estimate how much compression-induced error a surrogate model can tolerate without affecting its accuracy. Evaluation of two application simulations demonstrates that our approach significantly reduces memory/storage requirements and speeds up training while producing high-quality surrogate models. These results show that lossy compression saves data storage up to 23.7x and 39x with negligible impact on the quality of the surrogate model. Meanwhile, reducing the size of the training data set also enhances the data loading speed and reduces the training time by up to 3x.

Progress in AI has largely been driven by methods that assume less. As compute and data increase, approaches with weaker inductive biases generally outperform those with stronger assumptions. This is particularly characteristic of the field of Visual Representation Learning, where approaches have gone from being dominated by Supervised Learning, to Weakly Supervised Learning, to the now widespread success of Self-Supervised Learning without human labels. Yet, even modern Self-Supervised Learning approaches still depend on strong inductive biases such as augmentations, masking, or cropping. If this trend holds, even these remaining biases should become bottlenecks at scale – and our experiments confirm this: the optimal strength of inductive biases decreases as data grows. This motivates the search for approaches that rely on fewer assumptions. To this end, we introduce Temporal Difference in Vision (TDV), a new paradigm for self-supervised learning from video that avoids existing inductive biases, relying instead on a causal assumption that the past causes the future. TDV functions by jointly training an image encoder and a motion encoder so that the current frame's representation plus the encoded motion equals the next frame's representation. Despite not leveraging any strong inductive biases, TDV matches state-of-the-art recipes on dense spatial tasks, laying the foundation for representation learning without strong assumptions.

Agentic AI systems act through tools and sub-agents, yet the controls meant to bound their financial and environmental cost still sit on dashboards evaluated beside or after execution. Green SARC applies the SARC governance-by-architecture framework – four enforcement sites in the agent loop – to FinOps and GreenOps, contributing the theory of what to enforce and how to predict it. We report four policy-independent results. (i) The unconstrained "State Snowball" is $\Theta(n^2)$ in loop depth; on 3,000 real multi-step plans (SWE-rebench) it holds on 100%, with median curvature $\hat{c}_2=216$ exceeding the linear-accretion prediction $p/2=134$ – real plans accrete faster than the model. (ii) On real residuals the Normal-$\sigma$ gate under-covers (92% at nominal 95%); split-conformal calibration holds (95.2%). (iii) A soft Lagrangian penalty tuned to the budget in expectation breaches it on 91.5% of seeds; the architectural gate breaches 0%. (iv) Under binding budgets the gate's over-budget incidence is 0% on synthetic and real (BurstGPT) arrivals. End-to-end token/USD/carbon savings (47–55%) are real but policy-dependent in magnitude – set by a scope-cap knob, not by gate rejections. The library is open-source, dependency-free, and ships a regeneration script for every cited number.

Engineering LLM-native software remains a challenging and immature field. Current practice is largely exploratory, relying on experimentation and heuristic techniques such as prompting and context engineering. These, however, are low-level and lack the principled structure needed to support design-level reasoning or analysis. In contrast, traditional software engineering leverages modularity and abstraction to communicate and analyze system behavior. To bring similar rigor to LLM-native development, we propose methods for documenting generative flows and for stating properties of LLM-based software designs. Such methods must account for the stochastic, prompt-dependent behavior of large language models while remaining expressive enough to capture emergent phenomena. Our initial approach is based on graphical probabilistic models, tailored to capture phenomena characteristic of LLM-native systems. This framework – what we term Generation Networks – aims to provide a foundation for principled reasoning about generative interactions and system-level properties in LLM-centric software architectures.

Historical medical archives and traditional medicines hold immense potential for drug discovery and remain a primary source for current drug development. However, pre-ontological prose and idiosyncratic taxonomies prevent the standardization and medical modernization of the data for use in current biomedical pipelines. Furthermore, no existing LLM agent system, whether tool-calling, retrieval-augmented, or agentic deep-research, can convert such text into verifiable drug-discovery leads at scale. We close this gap with DeepRoot, a multi-agent LLM system that jointly builds and utilizes a verified knowledge graph, showing that grounding and reasoning – often conflated – are separable axes the system can compose for therapeutic reasoning. Applied to the Shen Nong Ben Cao Jing, DeepRoot recovers $10$ of $21$ held-out compound-disease treatment pairs at R@$20$ ($47.6\%$ vs $4.8\%$ for a raw corpus LLM and $\sim\!2.4\%$ random) and dominates an LLM-as-judge audit for reasoning quality over baseline LLMs and LLMs with direct tool-call access to the same APIs DeepRoot itself queries. Tool-using LLMs hallucinate evidence on $87\%$ of claims, versus 7-10% for DeepRoot. Graph-only inference hallucinates $0\%$ but ranks lowest on reasoning coherence; DeepRoot KG+LLM is the only condition to win on both axes, pointing toward a route for systematic mining and repurposing of historical medical knowledge.

Simulation is central to validating autonomous driving systems, yet current pipelines are limited by insufficient scenario diversity due to costly High Definition (HD) map creation. Scaling HD maps requires expensive data collection and manual processing. Moreover, existing generative models lack the fine-grained control necessary to target specific road topologies during generation. This paper presents a data-driven pipeline for controllable HD map generation using latent diffusion and ControlNet for spatial conditioning. To our knowledge, we are the first to inject spatial guidance signals into a diffusion model for HD map synthesis. Furthermore, our model supports adjustable conditioning strength through classifier-free guidance and city-level style transfer via city label conditioning. To complement existing metrics, we introduce two novel metrics to evaluate adherence to the control signal and similarity to ground-truth maps. Experiments demonstrate that our model generates realistic HD maps that faithfully follow input road topologies while accurately preserving city-specific details.

Cartesian Genetic Programming (CGP) is among the practical and popular forms of Genetic Programming as it uses a graph-based representation of programs. This paper presents a first runtime analysis of CGP in evolving Boolean functions using complete training sets. We prove an asymptotic bound $O(n D^5)$ for the expected number of fitness evaluations of CGP to construct a conjunction of $n$ inputs using at most $D \geq n-1$ binary gates, a minimal function set, and even with a strict survival selection. When the non-strict selection is used, the bound is improved to $O(n D^4)$. Our analysis reveals interesting characteristics of CGP induced search, which have been only observed empirically. In particular, enabling the acceptance of equally good solutions, including those with connected gates non-contributing to fitness, can lead to a speedup, and consequently a better asymptotic time bound. In contrast to conjunctions, we also prove a negative result which shows that CGP requires exponential time to evolve an exclusive disjunction. Experiments evolving conjunctions complement our theoretical findings. The use of incomplete training sets is found to further reduce the average number of fitness evaluations while maintaining a good level of generalisation.

Multi-turn agents that plan, invoke tools, and interact with environments offer a promising paradigm for solving complex tasks, yet their capabilities typically rely on very large models whose inference cost is prohibitive in practice. On-Policy Distillation (OPD) is a natural recipe for transferring such capabilities to smaller students, but we find that it suffers a characteristic failure mode in this setting: small student errors compound across turns and push the trajectory out of the teacher's familiar state distribution, so the teacher's supervision becomes least reliable precisely where the student needs it most. We propose Guided On-Policy Distillation (Guided-OPD), a simple yet effective algorithm that mixes teacher- and student-generated turns within each rollout and schedules the teacher's intervention probability along a curriculum that decays to zero. Strong guidance keeps early trajectories close to the teacher distribution and is then gradually withdrawn to recover the purely on-policy regime used at inference. On ALFWorld, ScienceWorld, and WebShop, distilling Qwen3 students from a Qwen3-30B-A3B teacher, Guided-OPD yields average relative gains of 21.1% in Score and 25.5% in Success Rate over vanilla OPD, with larger gains on smaller students.

Long-document multimodal question answering requires a system to locate sparse evidence in long PDFs and integrate clues from text, tables, images, charts, and complex layouts. Existing RAG methods mostly rely on fixed Top-k retrieval over text chunks or pages. Text retrieval can compress the context but often loses visual and layout information; page-level visual retrieval preserves the original page, yet it also sends large irrelevant regions to the reader, leading to a static trade-off among evidence coverage, noise, and inference cost. This paper proposes MAGE-RAG, a multigranular adaptive graph evidence framework for long-document multimodal QA. MAGE-RAG uses page retrieval as the entry point for query-time evidence construction. Offline, it builds an evidence graph with page nodes and element nodes, encoding containment, reading order, layout adjacency, section hierarchy, and semantic-neighbor relations. At query time, an online evidence controller iteratively activates, opens, searches, and prunes evidence under explicit budgets. The resulting evidence subgraph is then rendered into structured multimodal reader input, allowing the LVLM to consume compact and relevant evidence within a limited context. On LongDocURL and MMLongBench-Doc, we establish a unified comparison and analysis protocol covering Direct MLLM, Text RAG, Page-level Visual RAG, and Graph/Agentic RAG. Experiments show that MAGE-RAG achieves 52.75 overall accuracy on LongDocURL, and 53.26 accuracy with 51.19 F1 on MMLongBench-Doc. Fine-grained breakdowns, budget-performance curves, ablations, and trace-based analysis further show that query-time evidence subgraph construction can balance dispersed evidence coverage with context-noise control. Our code is available at https://github.com/laonuo2004/MAGE-RAG.git.

Where an LLM sits in an agent memory pipeline – between the recall plane that retrieves stored facts (extensively benchmarked) and the control plane that mutates them via supersede, release, purge (largely untested) – shapes which forgetting failure modes the system recovers. Comparing thirteen system configurations on a 385-case adversarial surface, we observe three placement regimes with partly complementary coverage: deterministic primitives suffice for lexical/temporal categories but fail canonicalization (5% on identifier-obfuscation, 0% on cross-lingual); inscribe-time LLM recovers canonicalization (100%) but cannot help intent-aware deletion (0% on prefix-collision and compound-fact); a mutation-time hook recovers intent-aware deletion (78-85%) and brightens nearly all categories simultaneously (91.7-93.2% overall, $0.17 per 385-case run, 2.3s/case mutation latency vs. 64-191ms/case deterministic, recall path unchanged). We expose the trade-off via ForgetEval, a 1000-case templated suite plus a 385-case adversarial layer (132 hand-crafted + 253 LLM-drafted oracle-validated) scored by deterministic substring match, paired with a six-method Adapter Protocol with honest N/A scoring that lets heterogeneous memory stores enter in 130 lines. Admission is corroborated by 10-annotator IAA (Fleiss' kappa = 0.958) and a 77-case external-authored subset (four blind contributors) that replicates the canonicalization asymmetry and amplifies the joint-placement lift (+27.8 pt). Production failures are predominantly forgetting failures rather than recall failures, yet existing benchmarks measure only recall. ForgetEval and all adapters are released under MIT.

Open-source LLM agent ecosystems are growing rapidly, yet the security of community-contributed skills - modular tool definitions that extend agent capabilities - remains largely unvetted. The gap we fill: existing scanners operate at the code layer and are structurally blind to instruction-layer and multi-agent risk - natural-language directives that hijack an agent, exfiltrate data through encoded side channels, or chain harm across pipelines - so what is needed is a semantic, multi-dimensional vetting system rather than another signature matcher. We present SKILLVETBENCH, a live public leaderboard on Hugging Face that uses an LLM-as-Judge to vet agent skills. What is new: SARS (Skill Agentic Risk Score), a five-dimensional agentic-risk metric with a principled weighted formula for instruction-following systems. What is integrated: full CVSS v4.0 vector decomposition and a ClawHub dual-view that places our LLM-generated review beside the official marketplace verdict. What is demonstrated: drawing on our companion benchmark paper [ 1], the LLM-as-Judge stage achieves zero false negatives across 78 confirmed-malicious skills and zero false positives across 22 benign controls, while the best static baseline (SKILLSIEVE) still misses 15%; for instruction-layer categories such as Prompt Injection and Memory Poisoning, conventional tools miss between 89% and 100% of threats (e.g., CODEBERT detects none of nine memory-poisoning skills). Detection rates vary from 35% to 95% across four LLM evaluators, motivating ensemble scoring in production deployments.

Flow matching is a powerful generative modeling framework, valued for its simplicity and strong empirical performance. However, its standard formulation treats signals on structured spaces, such as fMRI data on brain graphs, as points in Euclidean space, overlooking the rich topological features of their domains. To address this, we introduce topological flow matching, a topology-aware generalization of flow matching. We interpret flow matching as a framework for solving a degenerate Schrödinger bridge problem and inject topological information by augmenting the reference process with a Laplacian-derived drift. This principled modification captures the structure of the underlying domain while preserving the desirable properties of flow matching: a stable, simulation-free objective and deterministic sample paths. As a result, our framework serves as a drop-in replacement for standard flow matching. We demonstrate its effectiveness on diverse structured datasets, including brain fMRIs, ocean currents, seismic events, and traffic flows.

Understanding urban wellbeing from multimodal data requires integrating heterogeneous spatial and temporal signals, posing significant challenges for current multimodal large language models (MLLMs). We introduce UrbanWell, a large-scale benchmark designed to systematically evaluate the spatio-temporal reasoning capabilities of MLLMs for urban wellbeing analytics through joint modeling of satellite and street view imagery. UrbanWell spans 38 cities across multiple years and includes diverse indicators covering (1) environmental conditions (CO$_2$, NO$_2$, PM${2.5}$, and Normalized Difference Vegetation Index), (2) spatial accessibility (minimum distance to supermarkets and restaurants), (3) urban form (road length, road density, and land use), (4) urban vitality (population, economic activity diversity, and land use diversity), and (5) subjective perception attributes (e.g., safety, beauty, liveliness, wealth, and quietness). All indicators are aligned at grid level to enable standardized evaluation. Beyond static prediction, UrbanWell defines temporal reasoning tasks, including future value forecasting from historical observations and temporal trend classification. We benchmark 15 state-of-the-art representative MLLMs in a zero-shot setting, providing a comprehensive comparative evaluation across spatial and temporal dimensions. Experimental results indicate that while MLLMs capture salient spatial and perceptual cues, their performance varies substantially across heterogeneous urban indicators spanning environment and subjective perception. UrbanWell serves as a unified benchmark for evaluating multimodal spatial and temporal reasoning in urban wellbeing analytics, offering a standardized testbed for systematic assessment and future research on multimodal urban intelligence. Our codes and datasets are accessible via https://github.com/axin1301/UrbanWell-Benchmark.

Non-verbal vocalizations (NVs), such as laughter, sighs, and coughs, are important acoustic cues for emotion and intent. Existing speech quality assessment methods typically focus on overall naturalness, while non-verbal TTS evaluations mainly examine whether a target NV appears with the correct type and position. However, the perceptual quality of NV events themselves remains underexplored. To address this gap, we construct an NV-MOS dataset containing outputs from multiple NV-TTS systems and naturally occurring NV samples, with ratings collected from three acoustic experts on a perceptual quality scale. We further analyze audio-capable multimodal large language models such as Gemini and find clear inconsistencies between their scores and expert ratings. These results suggest that general-purpose multimodal models cannot reliably replace human judgments for NV quality assessment. We then propose NVMOS, to our knowledge the first model that can reliably predict the perceptual quality of NV events in speech. Experimental results show that, with a local NV-event focusing module, NVMOS reaches expert-level or stronger agreement with human MOS.

Large language model (LLM) systems are increasingly proposed to assist peer review, yet most evaluations judge the prose of machine-generated review text, not the validity of the numeric score a system assigns. We validate AIPR, which reads a submitted manuscript and emits five 0-100 quality dimensions and a weighted overall score, against the public decision outcomes of a major machine learning venue. AIPR grades by prompting alone, with no fine-tuning on reviews or decisions. Across 300 ICLR submissions with public decision tiers and reviewer ratings, graded under a frozen pipeline with hypotheses pre-registered before any score met any outcome, the overall score separates rejected from accepted submissions (AUROC 0.82, 95% CI 0.78-0.87), rises monotonically across tiers, and tracks the mean reviewer rating. The signal is strongest where we claim it: the lowest-scoring fifth is rejected far above the base rate, with oral papers absent. The validity comes mostly from the model: a one-paragraph prompt on the same model discriminates almost as well as the full pipeline (the small gap favours the pipeline but does not meet the pre-declared criterion, p = 0.09). What the engineering adds is reliability and a grounded review: AIPR's score barely moves across repeated runs (0.7 vs. 2.8 points within-paper SD) where the bare prompt swings, and the same pass returns a rubric-structured, evidence-grounded review rather than a bare number, with the human keeping the decision.