Papers

Language models hold latent quantities in a form they can report on, and more of a quantity is present in that form when the task requires reusing it flexibly. What causes a representation to enter that form is open, and the word workspace invites an admission story: a gate that decides what gets in. Testing it on open-weight models with Jacobian lenses, over a benchmark whose five arms share an identical context, we find no gate where it predicts one. Demand raises a concept's lens visibility beyond what applying an operator to a supplied value produces: +0.050 [+0.045, +0.057] in percentile rank on our primary checkpoint, positive on all four we measure, though that arm answers at ceiling and the accuracymatched contrast is stronger under that readout. At the same time one shared linear map decodes the variable from every arm, the control included, at 6.4-9.0x its selection-corrected floor. What produces the later readable form at the queried position is attention-mediated gathering inside a mid-depth window: separating patch depth from readout depth puts transport there at least 17x above anywhere shallower under non-saturating readouts, with no tested MLP output contributing positively inside it. Under the saturating percentile rank the same grid does not localise the window, which is a fact about that measure. An arm that needs the variable for nothing concentrates sevenfold less, so the window is demand-specific. That window has two measured edges, a survival failure below and destruction above, and it falls at the same fractional depth in a 64-layer hybrid and a 62-layer dense model from another family. We localise where the variable is installed and read, not the route from the passage, which transports nothing. But the readout is not a calibrated measure of use: three components move it to within 12% of one another and differ 7.4x in what they do to the answer.

Breast mass segmentation is an important step in computer-aided mammography, but it remains difficult because masses can have low contrast, irregular shapes, and boundaries that blend with surrounding breast tissue. To address this problem, we present DualMiT-Net, a dual-branch network that uses both a focused view of the mass and a wider view of the surrounding tissue. The local branch uses a Mix Transformer (MiT-B5) encoder to learn mass shape, texture, and boundary information, while the global branch uses an EfficientNet-B5 encoder to learn surrounding breast context. Features from the two branches are shared at the deeper encoder levels and are then progressively fused in a single decoder. A spatial gate controls how much global information is added during decoding. We also evaluated four input representations and selected a percentile-windowed mammogram combined with a Gabor texture response. The model was trained and evaluated on the mass subset of the Curated Breast Imaging Subset of the Digital Database for Screening Mammography (CBIS-DDSM) using a patient-level split. Across three training runs, DualMiT-Net with exponential moving average weights achieved a mean Dice coefficient of 0.9375 and a mean Intersection over Union of 0.8834. It also achieved better Dice and IoU scores than six standard encoder-decoder baselines trained using the same data and training settings. These results show that combining local mass information with wider breast context can provide accurate and consistent breast mass segmentation.

Deploying large language models (LLMs) for inference on edge devices is challenging due to severe memory and bandwidth constraints. While speculative decoding and Mixture-of-Experts (MoE) have been proposed to improve inference efficiency, naively combining them often incurs excessive verification overhead and poor expert reuse, limiting their effectiveness in memory-bound edge settings. In this work, we propose S2-MoE, an efficient self-speculative decoding framework for MoE inference on edge devices. S2-MoE reduces redundant verification through routing-aware adaptive speculative expansion, improves verification efficiency with reuse-aware expert gating, and aligns draft and target execution via shared context. Implemented in llama$.$cpp, S2-MoE achieves up to $5.3\times$ speedup (about $2.0\times$ on average) over standard autoregressive decoding across diverse MoE models and datasets on edge devices. Code is available at https://github.com/angerybob/S2-MoE.

Network incident response remains slow and labor-intensive as the defender must infer multi-stage attacks from partial observations and translate recovery decisions into reliable system commands. Decision-theoretic planners provide principled optimization but typically rely on abstract states and predefined actions, while large language model (LLM) agents can reason over operational context but may hallucinate attacks and responses. Toward automating response planning, we present a hierarchical agentic response framework that integrates LLM-based attack inference, rollout planning, and digital-twin validation. A fine-tuned LLM infers the attack progression and affected hosts from security alerts and system measurements. An emulated network digital twin replays the inferred attack and returns discrepancies between predicted and observed effects to calibrate the inference. A separately fine-tuned planning agent uses the rollout planning method to prioritize affected components at the tactical layer. At the operational layer, the planning agent proposes high-level recovery actions, and an execution agent translates selected actions into recovery and verification commands that are validated in the digital twin. We evaluate the framework on a 33-component enterprise-network testbed under three multi-stage attack scenarios. The results show that our framework outperforms frontier-LLM baselines in recovery success rate by 18–31%.

The rapid advancement of large language models (LLMs) has created a growing asymmetry in cybersecurity, where attack accelerates toward autonomous execution while defense remains predominantly human-intensive. Despite substantial prior work across cyber ranges, AI-driven attack, and AI-driven defense, this asymmetry persists. We trace it to a deeper root cause, that evolution itself has stalled on both sides at three layers. To overcome this, we propose co-evolution as the integrating insight, where attack and defense AI agents autonomously and safely drive each other's evolution through adversarial confrontation. Based on this insight, we present \sysevolve, comprising three co-designed components, \sysfield, \sysspear, and \sysarmor. \sysfield constructs realistic multi-host ranges. \sysspear generates efficient, safe attack schemes. \sysarmor performs real-time, interpretable defense. Together they form a self-driven adversarial loop restoring evolution at all three layers. In evaluation, \sysfield achieves zero-loss collection at 2.1% overhead and orchestrates 257 CVEs into 1,148 ranges, \sysspear improves attack success by over 25% over baseline LLMs, and \sysarmor achieves 10–1000$\times$ greater precision than prior systems and detects real APT attacks in production at Huawei and Sangfor. Our evaluation also reveals three findings about LLM agent capabilities. First, multi-step composition and larger topologies expose agent capability gaps hidden by single-step evaluations. Second, the bottleneck lies after initial access in post-compromise state utilization. Third, LLM agents are susceptible to environmental interference. When decoy endpoints are deployed in the range, agent timeouts triple and downstream completion disappears despite the success rates of initial accesses are unchanged.

Although visual reasoning is crucial for solving complex geometry tasks, existing vision-language models rely heavily on text-only reasoning. Some recent methods introduce intermediate visual states to facilitate reasoning, but they are often hindered by inaccurate geometric representations and low rendering fidelity, ultimately leading to unreliable outputs. To address these limitations, we propose MetaReason, a framework for multimodal reasoning in plane geometry that leverages structured meta-information to enable accurate auxiliary-line construction. The framework first parses geometric images into meta-information, performs controllable edits with predefined tools to synthesize high-fidelity visual states, and then conducts reasoning based on these augmented views. To support this framework, we construct TutorGeo, a comprehensive dataset containing 17k image-to-meta conversion samples, 60k text-only reasoning traces, and 60k interleaved multimodal reasoning traces. Using this dataset, we combine supervised fine-tuning and reinforcement learning to develop robust multimodal reasoning capabilities. We also introduce ExamGeo, a benchmark derived from real-world examination problems that enables systematic evaluation across varying difficulty levels. Experimental results demonstrate that MetaReason significantly outperforms existing open-source models and achieves competitive performance against proprietary models.

Lung cancer remains one of the leading causes of cancer-related mortality worldwide, and Computed Tomography (CT) is a primary imaging tool for screening and followup assessment. After pulmonary nodule detection, radiologists manually assess anatomical location, diameter, margin characteristics, and attenuation type to support risk assessment and clinical decision-making. However, this post-detection workflow is time-consuming and can be affected by inter-observer variability. Existing Artificial Intelligence methods often focus on isolated tasks, limiting their use as a unified, clinically grounded interpretation framework. This study presents FZ-VLM, a two-stage Florence-Zephyr Vision Language Model framework for unified structured pulmonary nodule characterization in lung CT. The framework uses a fine-tuned Florence-2 model to extract radiological attributes from expert-annotated 2D axial CT slices, while a Zephyr-7B model uses these attributes to generate nodule descriptions, follow-up recommendations, and longitudinal analyses. Results showed that the Stage 1 model achieved 77.18% accuracy for anatomical location, 67.96% accuracy for margin characteristics, and 79.13% accuracy for attenuation type, with a Mean Absolute Error of 2.58 mm for diameter estimation, outperforming evaluated GPT-4-based baselines as well as the human baseline. Expert radiologist evaluation of Stage 2 showed 93.9% accuracy, 98.6% completeness score, 76.1% clinical relevance, and an overall score of 89.5%. Safety analysis showed that most outputs were clinically safe, although some follow-up recommendations still required expert review. To the best of our knowledge, this study presents the first two-stage Vision-Language Model framework for structured nodule characterization and clinical decision-making.

Constructing special graphs is an important task within graph theory and computer science. Many popular graph constructions are the result of a comprehensive exploration of relevant graphs and human ingenuity. Given the rise of generative AI usage in mathematics, it is natural to test whether LLMs are able to construct graphs with specified properties using their reasoning capabilities. Unfortunately, many natural graph construction problems, such as finding extremal Ramsey-good graphs (i.e., avoiding specific monochromatic subgraphs), have been explored extensively in the literature, making it difficult to ascertain whether a construction is the product of an LLM's reasoning capabilities or its recollection from training data. In this work, we introduce RamseyGadgets, a novel dataset of 70 underexplored graph construction problems that require finding Ramsey-good graphs with special properties (e.g., containing an edge with a fixed color). These problems have reasonably sized solutions (at most 10 vertices) that can be verified by SAT solvers, making them suitable for automatic evaluation. Our dataset is easily expandable, as one can simply change the monochromatic subgraphs being avoided to obtain a new set of problems. We evaluate the performance of five open-source LLMs on our dataset and report the results. Our findings show that LLMs achieve only 37.70% accuracy on the hard-tier problems in our dataset, with Gemma-4-31B achieving the highest performance out of the five. We also showcase how our dataset allows us to ascertain what kind of hints help LLMs perform better at this task.

Quantum federated learning (QFL) lets multiple quantum clients collaboratively train quantum neural networks (QNNs) without sharing private local data. However, existing QFL methods commonly assume that all clients use the same ansatz, overlooking how heterogeneous client data affects ansatz suitability. Under class-imbalanced non-IID data, different clients may favor different ansatz structures, so a fixed ansatz can lead to unstable and unfair performance across clients. In this paper, we propose PAS-QFL, a Personalized Ansatz Selection framework for QFL under client data heterogeneity. Rather than treating the ansatz as a monolithic structure, PAS-QFL decomposes each client QNN into a globally shared ansatz and a client-specific private ansatz, and personalizes the structure of the private ansatz rather than only its parameters. The shared ansatz is placed first and selected by a stability-aware cross-client criterion so that its parameters can be reliably aggregated, while the private ansatz serves as a personalized decision head, selected per client by local Macro-F1 to adapt the shared representation to its local data. During training, each client updates both its shared and private parameters locally but uploads only the shared parameters, so federated aggregation stays well-defined while each client keeps its own private structure. PAS-QFL uses Macro-F1 as the primary selection metric to avoid misleading accuracy under class imbalance. Experiments on heterogeneous QFL tasks show that PAS-QFL improves average Macro-F1 over the existing fixed-ansatz QFL baselines, demonstrating the value of personalizing the ansatz structure for practical QFL.

Language-model systems increasingly read from stores they also write to, so a claim that was merely written earlier can return looking retrieved. We tested whether the message package carrying an unsupported assignment changes which answer a model gives in a synthetic lookup task. Claude Opus 5 selected a color code for a named item or abstained. In an exploratory four-arm study, false-code adoption was 0/24 with no target claim, 0/22 scorable trials when a prior assistant assertion named the target, 14/24 when a tool-result record named it, and 15/24 when that result used a ten-field metadata wrapper that marked it unchecked. The tool-result arm selected the record's code in 11/12 supported trials and 14/24 unsupported trials, ruling out a fixed output-token bias while leaving substantial planted-token heterogeneity. A document-preregistered replication reproduced the tool-result versus assistant-assertion gap, 7/24 against 0/24, one-sided Fisher exact p = 0.0047. The tool-result rate nevertheless fell from 14/24 to 7/24 across runs made four days apart. A second preregistered study gave the earlier comparison a live text control: both records were announced in advance and placed in the same final user turn, then target binding was swapped between the linked tool result and later inline JSON. Inline text was sufficient for false-code adoption in 60/60 trials; the tool-result condition produced 57/60, so the registered result-first superiority criterion failed, p = 1. The result does not show that tool results have no effect. It shows that native tool-result placement was not necessary and that this experiment did not find greater behavioral weight for the result package than for announced inline text. The findings concern a single model on one synthetic task template, accessed through one API.

Long-horizon robotic manipulation fundamentally relies on persistent spatial memory. However, existing 3D memory systems function merely as passive recorders: they store observations using fixed, hand-crafted rules, treating every scene element–whether a critical grasp target or an irrelevant background wall–with equal importance. In this paper, we propose a paradigm shift from passive storage to active, task-driven spatial memory. We argue that a robot's memory should not simply record what it sees, but actively learn how to remember–discovering which objects to track precisely, how aggressively to update them, and what to discard, all learned end-to-end without hand-designed rules. Crucially, this active paradigm is realized by unifying memory update and readout as two sides of the same cognitive process, enabling bidirectional flow where task needs shape update strategies and vice versa. To instantiate this vision, we introduce GaussMemory, which leverages 3D Gaussian Splatting as a persistent geometric substrate. On LIBERO, GaussMemory outperforms MemoryVLA on Goal and Long-10; on VLABench, it surpasses $\pi_0$-FAST by +5.2% (Track 1) and +6.0% (Track 6).

Transformers applied to spatial imperfect-information games must represent map geometry while tracking hidden entities through time. We ask whether geometry-aware positional encodings improve these capabilities, without claiming a new positional encoding. We construct a four-level benchmark on a hexagonal naval pursuit game: controlled geometry and topology probes, an exact-Bayes hidden-target tracking task, offline policy imitation at 1k and 10k games, and 7,200 fixed-seed games against three legacy opponents. Across matched Transformer backbones, HexRoPE reduces exact-belief posterior cross-entropy relative to no positional encoding by 0.278 on D6-transformed test orbits and 0.329 on a larger map; both hierarchical-bootstrap confidence intervals exclude zero, and both Holm-adjusted p-values are below 0.001. At 1k games, HexRoPE improves policy action accuracy by 4.63 percentage points over no encoding and 2.05 points over rectangular relative bias; the gains shrink to 1.55 and 0.41 points at 10k games. However, HexRoPE does not improve aggregate gameplay win rate: its paired effect over no encoding is -1.56 percentage points (95% CI [-4.50, 1.17]). Rectangular relative bias is strongest on D6 belief consistency but fails sharply when extrapolating from radius 3 to radius 4, while graph bias provides only a small blocked-edge gain. The results show that geometric inductive bias improves belief estimation and data-efficient imitation, but those representation gains do not automatically produce stronger closed-loop play.

This paper presents a demand-driven framework for on-demand Urban Air Mobility (UAM) network design that links vertiport siting, fleet simulation, and door-to-door travel-time feasibility. Demand is estimated from commuter and passenger activity data, converted into spatial trip-end points, and clustered using K-means to generate candidate vertiport locations. Candidate networks are screened using range and minimum station-spacing constraints, then evaluated with a discrete-event simulation that models multi-vehicle dispatch, deadhead relocation, battery swaps, and service regularity. Flight time and energy consumption are computed using a point-mass eVTOL performance model. In a Greater Los Angeles case study, the preferred design expands from four stations and four eVTOLs at low demand to sixteen stations and twelve eVTOLs at the highest tested demand level. Results show that larger fleets improve completion time and vehicle-arrival regularity but do not eliminate deadhead flights, indicating that spatial demand imbalance remains an operational burden. The travel-time savings analysis further suggests that UAM is most defensible for longer or congestion-heavy trips where sufficient non-flight time remains after accounting for flight time.

Pareto Conditioned Networks learn multiple multi-objective reinforcement learning behaviours by conditioning a single policy on a desired return command. However, the local mapping from command and state to action remains opaque. We propose command-space counterfactual explanations for PCNs: given a fixed state, original command, and foil action, we search, in a black-box setting, for a minimally changed desired-return command under which the same trained policy would choose the foil. Our contributions are threefold. First, we formulate PCN explanations as return-command interventions, using a return-only PCN variant that avoids the added ambiguity of horizon-conditioning. Second, we adapt adversarial machine learning methods to reinforcement-learning explanations. Third, we introduce a boundary-seeded directional search that improves over purely local optimization in the command-action landscape, resulting in our proposed approach CF-ZOO. The resulting explanations are actionable and intuitively expressed in the user's own preferences: "If your trade-off had shifted slightly towards X, the agent would have chosen Y."

Mesoscale eddies are fundamental to the ocean circulation, yet the extent to which submesoscale motions, a few kilometers across, influence mesoscale eddy energetics through a kinetic energy cascade remains uncertain. High-resolution simulations predict that submesoscale fronts are key regulators of the cascade, transferring energy both downscale towards dissipation and upscale to sustain and shape the seasonality of mesoscale eddies. Testing these predictions has remained difficult because existing observations and state estimates cannot resolve submesoscale currents over sufficiently broad domains. Here we map the ocean's submesoscale energy cascade by combining multi-source satellite observations with a generative deep learning framework, reconstructing gap-free, kilometer-scale surface currents with physically plausible dynamics learned from simulations. Applying this to the eddy-rich Agulhas Current system, we find that submesoscales energize the mesoscale through an upscale energy cascade above 10 km, contributing to the seasonality of mesoscale eddies. Below 10 km, convergence at submesoscale fronts drives a downscale cascade towards dissipation. Both upscale and downscale pathways concentrate within fronts, where cross-scale transfer is up to an order of magnitude more efficient. Despite their limited extent, fronts account for a substantial fraction of the domain-integrated cascade, establishing them as key regulators of the cascade and targets for next-generation eddy parameterizations.

Recent advances in Large Language Models (LLMs) have opened opportunities to apply high-level code transformations to the field of code optimization, and it has since emerged as one of the most fundamental tasks for LLMs to perform; however, at present, LLMs struggle to apply wide-ranging code optimization tasks due to both the complexity of the code and the inability to independently verify the correctness of the transformations. In this paper, we present the Trusted LLM (T-LLM) Compiler, which proposes an advancement in compiler technology through a collaborative effort involving high-level LLM code transformations, traditional compilers, and verification tools. Experimental results reveal that it can significantly improve code correctness when tested on a set of PolyBench/C benchmarks. Our approach facilitates iterative code optimization efforts with verification strategies that enable corrective actions. Through this approach, T-LLM Compiler achieves code optimization accuracy of up to 83.3% and a speedup of up to 16.1% on the PolyBench/C benchmarks, with the transformed code reaching an average of 26.7% speedup wrt standard baselines. Additionally, we release the project's source code to the open-source community.

Detecting opioid craving from wearable physiological signals is critical yet difficult, with the potential to support proactive interventions for individuals with opioid use disorder (OUD). This challenge is especially pronounced under subject-independent evaluation because craving is subjective, heterogeneous, and often physiologically entangled with stress. Our empirical analysis shows that stress elicits strong and reproducible autonomic responses, while craving-related signals are weaker, sparse, and largely embedded within stress-related physiology. We further show that psychological resilience, which shapes stress regulation and craving vulnerability, is not reliably observable from short-term wearable windows, but can be captured through reusable subject-level proxies, including post-stress heart-rate recovery and autobiographical memory recall. Motivated by these findings, we introduce RETRACE, a resilience-guided trait-conditioned framework for subject-independent craving estimation from wearable physiology. RETRACE reframes craving detection as trait-conditioned physiological interpretation: rather than assuming the same physiological pattern has the same meaning across individuals, it uses resilience-related subject context to guide inference. Technically, RETRACE introduces a novel dual-encoder design that separates generalizable stress physiology from subject-specific craving interpretation. It combines a frozen stress-pretrained encoder with resilience-conditioned craving encoder, using feature-level gating and representation-level fusion to enable lightweight personalization without target-user craving labels or per-user retraining. We evaluate RETRACE on a novel multimodal OUD dataset containing wearable physiology, stress and craving annotations, and autobiographical narratives. Under LOSO setup, RETRACE achieves up to 7% absolute improvement over the strongest baseline.

Bot accounts and coordinated influence operations are often discovered via heuristic methods, leaving a dearth of reliable ground-truth labels for training detection systems. To address this challenge, we study a natural question: can we generate labels to assist in learning embeddings in which bots and accounts from the same coordinated operation are close? We present FISSION, a method to generate labels by splitting each account's activity into positively labeled sub-accounts. Given this label source, we train detection models which preserve behavioral regularities recurring across positive sub-accounts. We evaluate FISSION and show it outperforms prior methods in detecting Wikipedia sockpuppets and Twitter/X bots.

A next-token probability says what a model predicts, not how much training support lies behind it. A Dirichlet head can represent this distinction by separating mean $m$ from concentration $S$, but decoupling does not identify what $S$ means. Here we propose an Evidential Next-Token Prediction (ENTOP) framework to audit this gap on character-level Moby-Dick, using exact 8-gram count as a reproducible lexical-support label and withholding count regression from 20% of context types. Standard implicit evidential training carries essentially no count signal beyond confidence on held-out-label types (partial Spearman $\rho = 0.001 \pm 0.014$), whereas explicit supervision generalizes ($\rho = 0.201 \pm 0.010$; matched-pair win $= 0.822 \pm 0.021$). CE predictive entropy is at chance for unseen 8-grams (AUROC $= 0.490 \pm 0.004$), while supervised vacuity reaches $0.772 \pm 0.003$, comparable with an indexed CE-representation baseline ($0.769$) but below the tautological corpus oracle ($1.000$). Neither longest-suffix nor representation-distance strata explain where amortization succeeds. Increasing count weight under the digamma objective improves support fit only by sacrificing prediction. A constant predictor wins natural log-RMSE, and vacuity does not improve error deferral. These results motivate a minimum evidence protocol—confidence control, matched pairs, held-out labels, a constant baseline, and a decision test—and show that concentration can pass identification while failing calibration and utility.

Accurately forecasting river runoff is key to managing water resources, controlling floods, and planning agriculture. This study examines the Ajichay River in northwest Iran, a major tributary of Lake Urmia that has experienced increasing water-related stress in recent years. We introduce a daily runoff prediction model based on Long Short-Term Memory (LSTM) networks. The model combines five LSTM units, each trained on different time intervals ranging from 2 to 6 days, to better capture variations in river flow patterns. To improve performance, each model was fine-tuned using Particle Swarm Optimization (PSO), a population-based optimization algorithm. The proposed approach was evaluated on unseen data from 2017-2018 using $R^2$, RMSE, and MSE as performance metrics. The results showed strong predictive accuracy, with $R^2$ values ranging from 74.95% to 91.42%. In addition, multiple feature-importance methods were applied to identify the most influential variables, providing further insight into the factors that drive runoff variations.

AI agents powered by large language models are evolving from isolated assistants into heterogeneous systems in which multiple agents, planners, tools, and execution backends operate over shared environments. In such settings, safety becomes a system-level action-governance problem: deciding whether a pending action should be committed given policy-relevant state accumulated across a session. Existing safeguards operate at fragmented boundaries, making it difficult to enforce shared policies over composed action flows across heterogeneous execution paths. We present OpenAgentFlow, a control-plane/action-plane architecture that establishes the action-commit boundary as a shared enforcement interface. GUI, API, tool, and LLM-generated actions are normalized into a common AgentEvent stream and mediated by a shared pre-execution Policy Enforcement Point, while provenance, session state, audit evidence, and updatable policies are maintained outside individual agents. This provides a common governance layer across incompatible executors and allows new policies to take effect without modifying agents, prompts, models, or execution paths. We evaluate OpenAgentFlow through complementary system evaluations spanning controlled action-flow tests, a public external benchmark, policy updates, and real Android execution. On a 300-case controlled suite, OpenAgentFlow achieves 94.00% accuracy and a 95.35% attack-block rate. On the complete 1,220-case AgentDojo-Traj split of TS-Bench, it achieves 97.62% accuracy, 96.59% unsafe-action recall, and a 1.96% safe false-intervention rate. New control-plane rules take effect without modifying protected agents, and the same enforcement path operates across live GUI, API/tool, and LLM-planned Android execution. These results show that a shared action-commit boundary provides a practical basis for system-wide governance across heterogeneous agent execution paths.

Large language models (LLMs) can generate plausible-sounding ETF portfolios while silently violating basic KYC-style constraints on risk, fees, and diversification. This is especially problematic in agentic multi-turn advisory systems, where each draft recommendation can become an action unless guarded by an auditable enforcement layer. We study a model-agnostic, asset-agnostic post-generation guardrail pipeline: (i) enforce a strict JSON allocation schema, (ii) validate allocations against numeric caps, and (iii) when violations occur, deterministically project the output to the nearest feasible portfolio via a convex quadratic program (QCQP). We introduce BiasMix-Finance (Mini), a compact stress-test benchmark for constrained decision-making under biased LLM generations, with a 16-ETF universe, three investor profiles, and eight bias prompts. Across three models and three inference modes (direct, critique, self-consistency), first-pass generations violate at least one cap in 47.6-85.7% of test cases (67.2% pooled), but the convex projection layer reduces final feasibility violations to 0% while requiring only a small correction distance (test pooled median D=||w*-w0||_2=0.066), indicating that the guardrail typically preserves the intent of the original allocation. We report violation rates and correction distances with confidence intervals, and paired model comparisons with multiple-testing correction. To support reproducibility, we release the dataset, prompts, caps, and code in our public GitHub repository.

Graph RAG connects facts no single passage states, but implementations pay three times: in infrastructure, keeping vector store, graph database and document store in sync; in quality, because a pipeline that never refuses extractor output stores edges asserting nothing; and over time, because a graph that only accumulates treats superseded and current facts alike. post-graph-rag is an open-source engine addressing all three. Chunks with embeddings, a canonical entity graph and community summaries live in one PostgreSQL database, with pgvector for search and edge tables for traversal. Extraction output is validated before writing: vague predicates, pronominal names and bare quantities are rejected, predicates normalise onto an optional vocabulary, and entities resolve to one vertex per canonical name. A bi-temporal layer records when a relation held and when the system believed it, superseding incompatible earlier assertions from document order. Against LightRAG on three corpora with extraction and embedding models fixed, it builds a denser and more queryable graph, and supersedes relationships where a baseline with no temporal model supersedes none. On LongMemEval, 500 questions of long-horizon chat memory, it scores 94.0 percent with gemini-3.6-flash against 71.2 for Zep's gpt-4o and 60.2 for full context, leading on all six question types. The largest single contribution is temporal grounding in the prompt: carrying each relation's validity period through to synthesis moves temporal reasoning from 0.496 to 0.881, ablated paired on one graph per instance. On ECT-QA, earnings-call transcripts restating every metric each quarter, it scores 0.807 under that benchmark's own protocol against 0.599, 0.406 and 0.405 published for TG-RAG, LightRAG and GraphRAG. Code: post-graph-rag https://github.com/crajah/post-graph-rag; post-graph https://github.com/crajah/post-graph

On-policy self-distillation (OPSD) gives language agents dense token-level supervision from a privileged self-teacher on the policy's own trajectories. Existing methods allocate this supervision mainly by teacher trust, but trust does not reveal whether emphasizing a token supports the current policy objective. We call this the trust-utility mismatch and introduce Influence Calibration for Self-Distillation (ICSD). For each supervised token, ICSD measures the first-order response of its importance-weighted RL surrogate contribution to a teacher-directed output perturbation. Batch-adaptive calibration converts this non-stationary signal into a bounded allocation weight while preserving the original auxiliary-loss mass within each action turn. These detached weights affect only the distillation loss and require no additional model pass. Across ALFWorld, WebShop, and Search-QA, ICSD improves all matched aggregate metrics over trust-only allocation under Group Relative Policy Optimization (GRPO) and Group-in-Group Policy Optimization (GiGPO), across two model families spanning 1.5B to 7B. At 7B, it reaches 96.1% ALFWorld success and a WebShop score of 93.1. Frozen-batch analyses show that ICSD reduces teacher-supported mass assigned to objective-opposed tokens from 60.1% to 37.8% and raises cosine compatibility with the RL gradient by 0.192. A companion repository is avail- able at https://github.com/lanqz7766/Influence-Calibration-for-On-Policy-Self-Distillation-in-Agentic-RL.

Agent skills are often injected in full on every request, increasing token cost. We compare four content-preserving loading methods: Full, Skill Block, Reference, and Hybrid. Across SearchQA, SpreadsheetBench, ALFWorld, ScienceWorld, and SynthProc, we measure token usage using raw input for single-turn tasks and cache-correct effective input for multi-turn tasks. Results show no universal winner. Hybrid reduces input by 27.4% on SearchQA and 39.8% on SpreadsheetBench. On large multi-turn skills, Skill Block and Hybrid achieve substantial reductions, reaching 62.5% and 52.8% on ScienceWorld and 73.0% and 66.6% on SynthProc. ALFWorld shows smaller gains because procedures are short and repeatedly needed. Paired outcome tests detect no quality differences, though they do not establish equivalence. Overall, conditional loading is most beneficial when large portions of a skill are not needed on every turn.

This paper presents the first known empirical investigation of annotator and reviewer performance across multi-source remotely sensed imagery, evaluating human labeling across drone, crewed aviation, and satellite views. Because existing aerial imagery datasets rely predominantly on single-source imagery, there is no currently established state of practice for efficiently allocating human labor to curate large-scale, multi-source aerial datasets. This work addresses this limitation by analyzing annotator and reviewer performance within a post-disaster building damage assessment dataset of 9 disasters, where 20041 buildings in drone, 20695 buildings in crewed aviation, and 33392 buildings in satellite imagery were labeled. These labels, provided by 187 annotators, were then refined through two successive quality-control stages: a single-reviewer pass followed by a consensus-committee review. Our analysis reveals two findings that raise questions for standard crowd-sourcing practices. First, initial annotations were revised by the final committee at rates that rise steeply from higher- to lower-resolution sources (25.27% for crewed aviation and 36.95% for satellite), with the same ordering at every observed workflow stage. Second, a single individual review reduced but did not resolve this disagreement: after review, the committee still revised 6.85% of drone, 14.05% of crewed, and 20.86% of satellite labels. These observations suggest that, in workflows like this one, uniform review allocation leaves the most residual disagreement in lower-resolution imagery. Based on this evidence, and consistent with prior work on adaptive task assignment and budget-aware quality control, this paper offers three recommendations for multi-source dataset curation.

Agent evaluations commonly score the state observed when a run stops and count the run as one trial. Interpreting that score as a final result from a separate trial requires outcome finality and cross-unit separation. Outcome finality requires that later events cannot change the claimed result, while cross-unit separation requires that earlier runs cannot change the relevant conditions of later ones. The endpoint establishes neither condition by itself, and the two can hold independently. Waiting for a delayed outcome may settle the label even though its state remains available to another run. Isolation may prevent carryover even though the scored outcome remains unresolved. We develop a completion argument that identifies the evidence needed for each decision. A final success or failure label is justified only when every relevant effect is resolved or bounded tightly enough to fix the outcome. Any remaining uncertainty must be reported. First, in a controlled replay with fixed agent actions, we find that endpoint and terminal labels differ for every nonzero-delay operation and that a delayed write changes the next run's score under shared state but has no such effect after namespacing or verified reset. Second, in a review of ten public protocols, we find that reset or deliberate retention is documented explicitly more often than unfinished operations or evidence for separate scoring. Finally, we propose an open-effects record for operations and resources that may remain relevant after the endpoint, their status, and their possible effects on the scored outcome or another run.

Small proxy models are commonly used to identify data mixtures for larger-scale training. We ask whether their training trajectories reveal another transferable structure: the order in which larger models should resolve skill bottlenecks. We formulate first-passage skill training, where each monitored skill has a target floor and the objective is to minimize the tokens required to reach all floors. We introduce LogFloor, a closed-loop controller that directs each round toward current bottlenecks, producing phase-ordered resolution trajectories. Across five bAbI skill slices on Qwen2.5-1.5B, LogFloor reduces token cost by 56.2% on average. In 70M-to-12B transfer, three-round replay of a 70M scout path reaches every floor in all eight target runs, saving 30.9% by pair mean, 39.4% in pooled training tokens, and 37.6% under source-cost accounting. On MMLU-control, a frozen scout path succeeds across all eight 12B runs. Collapsing a path to its static marginal mixture or reversing its phase order removes most benefits, while bottleneck labels alone remain partially useful. These results identify phase-ordered bottleneck resolution as a transferable curriculum structure for monitored skill-targeted training.

Multi-agent large language model (LLM) systems can improve reasoning by spending more computation, but deployment requires deciding when extra collaboration is worth its cost. We isolate this decision by running every problem under four protocols while holding the solver fixed within each setting: direct solving (Baseline), iterative self-correction (Single), planner-executor-reviewer collaboration (PER), and multi-agent deliberation (Broadcast). The primary benchmark comprises 4,181 competition-level math problems; paired robustness checks cover four benchmarks spanning competition math, biology, and broader science with two solver families. Across fixed policies, trained routers, and frozen LLM routers, conservative policies under-escalate, whereas higher-solve frozen routers often over-escalate. A post-answer, pre-collaboration gpt-oss-120b probe ranks Baseline failures with 0.8847 AUROC (4,151 parseable cases; 95% CI [0.8732, 0.8955]). The same score remains informative for predicting whether any collaboration helps (0.7683 AUPRC), but is much weaker for identifying PER- or Broadcast-specific value (0.1674 and 0.1041 AUPRC). Separately, the pre-answer self-confidence gate reaches 78.0% solve at 45K tokens, compared with 73.8% at 71.3K for a frozen gpt-oss-120b router and 92.4% for a retrospective fixed-order oracle. Across 10 paired model-condition settings, the oracle adds 23.2-58.3 points of retrospective coverage over Baseline, but protocol profiles vary by task. In the six settings with held-out router evaluations, oracle gaps remain 18.5-28.9 points. Confidence can therefore support initial escalation, while protocol-specific cost-aware routing remains unresolved.

Spatial transcriptomics (ST) links tissue morphology with molecular programs, motivating multimodal pretraining methods that align histology images with gene expression. However, existing approaches suffer from two key limitations: spatially informative gene selection is often dominated by ubiquitous housekeeping genes, leading to weakly discriminative representations, and independent spot-patch alignment fails to capture spatial dependencies that are critical for tissue organization. To address these challenges, we introduce PaSTel, a hierarchical multimodal pretraining framework that integrates biological priors at three levels. At the spot level, TF-IDF reweighting is used to identify spatially informative genes; at the functional level, curated KEGG pathways serve as anchors for encoding global biological semantics; and at the regional level, spatial clustering aggregates neighboring spots to model meso-scale tissue structure. Across multiple downstream tasks, PaSTel consistently outperforms existing vision and vision-omics encoders, demonstrating that incorporating multiscale biological priors yields more informative and transferable representations for spatial transcriptomics.