Papers

This paper describes DS@GT ARC's third-place solution to the PlantCLEF 2026 challenge on multi-species plant identification in vegetation quadrat images, where systems must predict every species present in high-resolution (~3000 x 3000 pixel) plot photographs while training only on single-label images of individual plants. The pipeline is built around a fine-tuned DINOv2 ViT-L/14 classifier applied over a multi-scale tile decomposition of each quadrat, with per-tile predictions blended with a FAISS kNN retriever and post-processed by source-aware temporal fusion across repeated plot visits, a habitat-fit demotion that injects geographic and altitude priors from the training data, and a South-Western Europe geographic mask. Habitat-fit demotion and multi-scale aggregation are the largest individual contributors in the ablations. Two complementary training-centric directions, a cross-region transformer with noisy-student distillation on the LUCAS dataset and a label-as-query transformer decoder over synthetic CLS-domain pseudo-quadrats, yielded null results. An inference-time augmentation with instance-aware segmentation crops also did not improve performance. The selected submission reaches a private-leaderboard macro-F1 of 0.43902 (third place; public 0.51096); an unselected configuration of the same pipeline scored above 0.45 on the private set. Code: https://github.com/dsgt-arc/plantclef-2026.

While reinforcement learning with verifiable rewards (RLVR) is widely used to improve the reasoning capabilities of large language models (LLMs), the generalizability of the resulting models remains poorly understood. In this work, we establish the first non-vacuous generalization bounds for parameter-efficient RLVR fine-tuning at the billion-parameter scale. Our approach adapts PAC-Bayes compression bounds to this setting, and addresses the inherent stochasticity of token generation by applying the Gumbel-max reparameterization trick. To operationalize these bounds, we propose the Progressive RLVR framework, which integrates RLVR with on-policy distillation, TinyLoRA, and model quantization. Progressive RLVR empirically retains 84-97% performance of standard LoRA fine-tuning while producing models that are 14,796x more compressible. We show that this framework yields non-vacuous generalization bounds in four domains: mathematical problem-solving, programming, general-knowledge reasoning, and Text-to-SQL. Our bounds exceed the accuracy of the base model by 9-51% and lie within 6-11% of the accuracy of the fine-tuned models.

Multi-modal Large Language Models (MLLMs) have made substantial advances on benchmarks, yet their real-world effectiveness remains uncertain. This gap stems from the fundamental misalignment between benchmarks in controlled, static settings and the dynamic, interactive, and contextualized nature of real-world applications. To bridge this gap, we propose CEDI (Contextualized Evaluations of MLLMs through Dynamic, multi-round Interactions), a framework that recasts evaluation as a three-party interaction between an evaluatee model, an automated examiner, and a grader. The examiner conducts multi-turn, semi-structured conversation guided by a graph-based representation of the task. By navigating state-space transitions, CEDI deploys diverse strategies, from clarification requests to adversarial probes, to elicit performance evidence. We apply CEDI to visual hallucinations. Empirical results across multiple models, diverse settings, datasets, and domains show that contextualized, interactive evaluations reveal not only significantly more hallucinations than conventional static evaluation but also ones that more closely resemble those arising in practical use cases. We further show that hallucinations often accumulate over long contexts, through self-reinforcing dialogue history, and models are particularly vulnerable to questions requiring premise rejection or refusal. Together, these findings highlight CEDI as a step toward realistic, systematic, and ecologically valid assessments of MLLMs' capabilities. Code is available at github.com/williamium3000/cedi.

Complex knowledge base question answering (KBQA) is commonly approached through either information retrieval over a question-specific subgraph or semantic parsing into an executable logical form. We study the latter paradigm. Recent large language model agents make semantic parsing interactive: they alternate between reasoning, querying the knowledge base, and extending a partial SPARQL query. This interleaving reduces reliance on one-shot generation, but makes the quality of KB grounding depend on what the interaction tools expose. Existing agents retrieve or prune candidate properties mainly through lexical relevance and instance-level observations, without systematically conditioning on entity types, property domains and ranges, or the expected answer type. We call this failure mode type-blind grounding. It enlarges the grounding search space and often produces plausible-looking but semantically incompatible triple patterns that execute to empty results. We propose SAGA (\underline{S}chema-\underline{A}ware \underline{G}rounding for \underline{A}gentic Text-to-SPARQL Generation), a training-free framework that turns property exploration into a schema-constrained grounding operation. SAGA maintains a persistent bidirectional type state, filters known-incompatible property candidates at construction time, presents the remaining graph patterns in a compact schema-annotated format, and handles missing schema information permissively through empirical and trace-local evidence. Across nine benchmark settings over Wikidata and Freebase, SAGA achieves the highest F1 on all nine settings and the highest exact-match accuracy on eight, while reducing empty-result queries across all reported Wikidata settings.

Edge computing brings unique challenges as the resources on the edge are highly diverse in capabilities and capacities, and highly distributed across many users and the physical world. Existing distributed computing frameworks cannot adequately handle this level of heterogeneity and distribution. This paper proposes EdgeFaaS, a novel function-based edge computing framework to enable edge applications to effectively utilize heterogeneous resources distributed across the Internet of Things (IoT), edge, and cloud for computing. It proposes function virtualization and storage virtualization to abstract distributed and heterogeneous physical resources and provides consistent virtual interfaces for deploying and executing functions and storing and accessing data. EdgeFaaS provides comprehensive support to diverse edge computing workflows, and at the same time allows users to flexibly adjust the configurations and explore various important tradeoffs. To demonstrate its usability, the paper also presents the implementation and evaluation of three representative workflows on EdgeFaaS for video analytics, federated learning, and audio classification, on a real testbed of 100+ geographically distributed IoT devices, edge servers, and cloud services. EdgeFaaS allows users to flexibly explore the deployment configurations of these workflows over distributed and heterogeneous resources. For example, users can easily vary the function placement of the video processing pipeline across IoT, edge, and cloud resources and study the tradeoff between computation and communication costs; users can also flexibly adjust the cluster count and size in the hierarchical federated learning system and explore the tradeoff between training accuracy and speed.

Large language model (LLM)-based generative agents simulate human behavior through long-horizon decision-making processes that comprise intermediate steps such as planning, memory retrieval, reflection, and action selection. However, fine-grained human annotations of these intermediate steps remain scarce, and existing agents are not grounded in human preferences over such intermediate decisions. To address this gap, we introduce \method, an interactive simulation interface that enables us to collect step-level human preference supervision over agent decision trajectories, leading to a dataset of 57K fine-grained annotations. We conduct step-level preference learning on open-weight language models using supervised finetuning and direct preference optimization on this data, consistently improving simulation fidelity, coordination, and interaction quality, and inducing more socially effective agent behavior. Our results show that step-level human supervision is an effective training signal for improving both local decision quality and long-horizon agent behavior.

This paper details the DS@GT ARC team's approach to BirdCLEF+ 2026, multi-label detection of animal vocalizations in soundscapes from the Pantanal wetlands. The 2026 edition adds about an hour of labeled soundscapes, shifting the task toward supervised pipelines fit to the labeled set. First, we build a competitive supervised baseline that ensembles a frozen Perch v2 backbone, a trained HGNetV2-B0 sound-event-detection network, and a non-bird prototypical head, reaching a private leaderboard score of 0.936 at rank 1894 within a 90-minute CPU budget. Second, we ask whether token-based representations can compete, contrasting codec representations from neural audio codecs against semantic representations from foundational embeddings. We compare two bioacoustic specialist models against four token-based encoders trained on AudioSet. The repository for this work can be found at https://github.com/dsgt-arc/birdclef-2026.

Large Language Models (LLMs) have accelerated the adoption of software development agents, now widely available as Integrated Development Environment (IDE) extensions and standalone applications. While these agents are typically general-purpose, it remains unclear whether specialist agents justify their additional development effort. We investigate this question in the context of business process automation, focusing on the transformation of Business Process Model and Notation (BPMN) diagrams into executable agentic workflows. Since BPMN specifies explicit control-flow semantics, we focus on deterministic workflows in which a fixed process model and inputs uniquely determine the executed path. We introduce a specialist workflow for this task and compare it against generalist agents such as Roo and Cline. Our results show that the specialist solution produces agents that outperform generalist baselines by approximately 9-20 percentage points in tool-use exactness, 2-4x in penalty-adjusted latency, and 3x fewer tool-call errors, while reducing generation token cost by over 95% and eliminating repair iterations. We also find that generalist agents generate code inconsistently in both functionality and quality, limiting their suitability for industrial settings where reliability and maintainability are essential.

Computer-use agents are becoming capable software operators, but their interface to desktop applications is still often a brittle motor layer: they look at screenshots, predict coordinates, click, and hope that the visible state changed as intended. This collapses target grounding, action execution, and outcome verification into a single ambiguous operation. We present Tactile, an open-source tool layer that gives agents a more reliable "hands and feet" for desktop use. Tactile converts heterogeneous UI evidence–operating-system accessibility semantics, OCR-grounded text, and visual fallback regions–into action-grounded interface states: compact target candidates with source labels, roles or text, state, geometry, executable affordances, and verification cues. Agents operate through an observe-ground-act-verify loop that prefers native semantic actions when available, falls back to OCR-grounded coordinates when visible text is the best evidence, and keeps full provenance for replay and failure attribution. On macOSWorld-style tasks, adding Tactile improves Codex Success@100 from 41.1% to 50.0% overall and from 45.2% to 55.3% on accessibility-adapted tasks; a 96-task cross-agent subset shows consistent gains across Codex, Claude Code, OpenCode, and Goose. These results suggest that reliable computer use requires not only stronger models, but also a reusable execution substrate that exposes software actions as semantic, verifiable, and auditable objects rather than anonymous screen coordinates.

This study investigated the public acceptance of Society of Automotive Engineers Level 3 conditionally automated cars, which can self-drive under certain specified conditions but require the human driver to remain ready to resume control when requested. Previous Unified Theory of Acceptance and Use of Technology 2 (UTAUT2)-based research has focused mainly on European samples, and so it is still unclear whether the same factors shape acceptance across broader world regions. This knowledge gap was addressed using the L3Pilot Global User Acceptance Survey. From an original dataset of 18,631 respondents, the final analytic sample comprised 18,603 respondents from 17 countries across Africa, Asia, Europe, North America, and South America. The data were analyzed using a UTAUT2-based structural equation model to examine how performance expectancy, effort expectancy, social influence, facilitating conditions, and hedonic motivation shape the intention to use Level 3 cars. The model showed strong explanatory power. Across the analytic sample, the intention to use Level 3 cars was driven mainly by performance expectancy, social influence, and hedonic motivation. Effort expectancy and facilitating conditions also contributed, but they played smaller direct roles. Age, gender, and previous experience with advanced driver assistance systems were statistically significant, but comparatively weak predictors. Overall, the findings suggest that the acceptance of Level 3 automated cars depends less on demographic characteristics or ease-of-use concerns and more on whether people see the technology as useful, socially supported, and enjoyable to use.

The natural way to review a long recording or document with a multimodal model is to hand it the raw source and ask for a review in one call. We show that this quietly fails: the model satisfices, dropping roughly a third of the content and embellishing the rest. The failure is not perception–almost all of the dropped content reappears when the same model is simply asked to transcribe the source. The bottleneck is generation under load: a single pass cannot perceive, reason over, and write a long faithful review at the same time, because doing all three competes for one output. We rule out the obvious alternatives. It is not the modality: models read text and an image of the same text equally well. And it is not merely a matter of thinking harder: giving the single pass a far larger reasoning budget does not recover the lost content, because the model spends that budget planning a review rather than writing the source down. What works is to split the labor across two same-weights passes–first transcribe, then review the transcript–so each step gets a full output budget of its own. This transcribe-then-review decomposition improves both faithfulness and coverage across a 21-source suite. The benefit is not uniform: we observe that it helps most where the one-pass baseline is weakest and little where that baseline is already strong, a pattern that also tracks the source's length and modality. Decomposition comes with two failure modes–the review pass running out of room on very long sources, and confabulating from memory once the grounding source is removed.

Early detection of dental caries remains challenging due to limitations in traditional diagnostic methods, particularly for proximal lesions in posterior teeth. Deep learning models show promise for automated caries detection but face scalability constraints due to requirements for large volumes of expertly annotated training data. This study presents a dual-stage deep learning framework combining Faster R-CNN for tooth localization with U-Net for pixel-wise caries segmentation in panoramic radiographs. We developed a systematic transformation pipeline to convert large-scale polygon-annotated datasets into high-resolution binary segmentation masks, enabling pixel-wise supervised learning. The framework was trained using both expert-verified datasets and algorithmically processed labels from 3,000 panoramic images. Our approach achieved robust performance with an IoU of 0.9013, Dice coefficient of 0.9482, Recall of 0.9433, and Precision of 0.9774, demonstrating superior accuracy compared to existing methods while significantly reducing false-positive rates. The dual-stage framework effectively addresses data annotation bottlenecks in dental AI applications and demonstrates potential for scalable, automated caries detection systems that can improve diagnostic consistency and support clinical decision-making.

Reducing radiation dose in computed tomography significantly degrades image quality and poses challenges for accurate and clinically reliable reconstruction. While recent approaches have shown promise for low-dose CT, they often struggle to generalize across continuous and previously unseen dose levels, leading to artifacts and loss of anatomical detail. To address these limitations, we propose Dose-Aware Cold Diffusion (DACD), a physics-consistent reconstruction framework that explicitly models radiation dose as a continuous latent factor within a cold diffusion process. The proposed DACD framework integrates image-based dose-aware perception, multi-scale structural prior extraction, and dose-calibrated step allocation to adaptively guide the denoising trajectory. In addition, an iterative forward-backprojection correction is incorporated into the reverse refinement process to enforce projection-domain data consistency. Extensive experiments on three public benchmarks, including Mayo-2020, Mayo-2016, and LoDoPaB-CT, demonstrate that DACD consistently outperforms state-of-the-art diffusion-based and physics-guided methods in both quantitative accuracy and visual fidelity, particularly under ultra-low-dose conditions. The results show that DACD achieves robust generalization across a continuous range of dose levels, including those unseen during training.

Existing LLM moral evaluations typically present models with isolated moral vignettes and elicit a single-shot decision, neglecting a factor known to profoundly influence human moral behavior: consequence feedback. We introduce TPvG (Text-based Pain-versus-Gain), adapted from a human moral paradigm, which embeds consequence feedback into an everyday moral dilemma of not harming others versus maximising self-gain. TPvG comprises five moral decision tasks, progressing from minimal-context one-shot choices to sequential decisions with explicit consequence feedback. Our results show that LLM moral decisions were strongly affected by decision format (one-shot versus sequential), and explicit receiver feedback produced heterogeneous effects across models. Furthermore, LLM responses to explicit receiver feedback diverged from the human reference pattern, suggesting potentially different decision processes. These findings highlight the need to evaluate whether LLM moral behavior remains stable in high-stakes interactive settings.

As large language models (LLMs) are increasingly used for everyday decision-making advice, whether a model shifts the direction of its advice according to the user's emotional state has become an important safety problem. We test whether emotional expression increases a model's endorsement (encouragement to proceed) when a user, holding the same objective information, is overconfident about a premature decision (e.g., quitting a stable job on weak evidence). As a key control, we include a no-emotion multi-turn (neutral) condition that holds factual content and the number of conversational turns constant, isolating the effect of emotion from that of conversation length. We exposed six commercial models (top-tier and mid-tier models from OpenAI, Anthropic, and Google) to three scenarios (career change, business expansion, emigration) across three conditions (cold/neutral/distress) with six repetitions each, yielding 324 conversations, and measured endorsement strength (0-100) via an eight-item rubric-based automated scoring. Emotional expression significantly increased endorsement (neutral 18.6 to distress 31.5, +12.9 points; mixed-effects $\beta = +12.9$, $p < .001$; Cohen's d = 0.51), and this was not explained by conversation length (cold-neutral difference non-significant, $p = .083$). Critically, the vulnerability varied by individual model rather than by price tier: five of six models showed a significant emotion effect, including the top-tier flagships Gemini 3.1 Pro and GPT-5.5, while only Claude Opus showed no significant change. Results were reproduced with an independent non-Google judge model ($\rho = .89$) and agreed in rank with two human coders ($\rho = .70$). Through a controlled design that separates emotion from conversational context, we show that emotional context increases LLM sycophancy even in top-tier flagship models.

This position paper argues that human-centered explainable AI (HCXAI) should incorporate insights from the psychology of information seeking. Drawing on Sharot and Sunstein's framework of information-seeking motives, we propose that people evaluate whether to engage with explanations based on three types of expected utility: instrumental (will it help me act better?), hedonic (will it make me feel better?), and cognitive (will it improve my understanding?). Each utility is estimated through a lens shaped by well-documented cognitive biases, including illusion of control, automation bias, unrealistic optimism, impact bias, overconfidence, and confirmation bias. These biases can lead to two failure modes: excessive information-seeking that fragments attention without improving decisions, and insufficient information-seeking that leaves critical risks and misunderstandings unexamined. This challenge is particularly acute for agentic AI systems, where explanations must support not just understanding a single output but anticipating cascading actions, assessing risks, and deciding when to intervene. By integrating information-seeking psychology into HCXAI, we advocate for a shift from making explanations available to making them sought: designing systems that account for when and why users actually want to know.

Student burnout is highly prevalent in higher education, with reported rates ranging from 12% to over 70% and consistently exceeding those of the working population - yet it is typically identified only retrospectively, after academic decline has already occurred. A contributing factor is that students have little structured visibility into their own study behaviour, and existing productivity tools record activity without interpreting it. This paper presents RIACT (Record, Insight, Analyze, Coach, Track), a web-based application that combines structured study session logging with a hybrid AI architecture to surface personalized insights and early burnout signals. Students log sessions by location and time; the system computes net focus time by accounting for breaks, detects burnout signals through transparent, deterministic rules operating on week-over-week behavioural comparisons, and uses a large language model - constrained to a fixed output schema - to contextualize patterns and generate personalized recommendations. The design embeds responsible AI principles throughout: warnings are governed by auditable rules rather than model judgement, all output is framed as an observation rather than a diagnosis and data collection is limited to self-logged behavioural fields. We describe the system's design rationale, situate it within the literature on student burnout and explainable AI in education and propose an evaluation framework for validating its behavioural signals against established burnout instruments.

High-confidence errors in large language models are often treated as evidence of fragile internal inference. We study a different possibility: stable miscalibration, where a confident wrong answer remains locally stable under small perturbations. We combine two diagnostics: a label-aware output-level audit score that ranks domains by confidence variation and overconfident mistakes under a forced-answer baseline, and an internal sensitivity probe that measures hidden-state movement. On a multi-domain binary factual audit set, this audit score tracks where abstention-aware self-critique reduces decision loss, although direct labeled baselines rank the same gain more strongly. Internally, self-critical prompting consistently reduces hidden-state sensitivity across layers in three open-weight models. This supports prompt-induced local stabilization rather than a purely output-level abstention pattern, but it does not imply calibration: audit-defined overconfident errors are not clearly more locally sensitive than confidently correct answers, so some high-confidence errors may be stable and miscalibrated rather than simply fragile.

Hallucination remains a persistent challenge for Multimodal Large Language Models (MLLMs), severely limiting their reliability in high-stakes applications. Existing evaluations, predominantly based on static benchmarks, suffer from narrow taxonomical coverage and rapid performance saturation, failing to reflect model robustness in evolving real-world scenarios. To bridge this gap, we present a systematic evaluation framework integrating a comprehensive benchmark with self-evolving stress testing. First, we introduce UniHall, a fine-grained dataset grounded in a unified taxonomy spanning Object, Instruction, and Knowledge dimensions. Second, to address benchmark saturation, we propose Self-Adaptive Multimodal Fuzzing (SAMF), a self-adaptive framework that employs evolutionary mutation strategies to explore the boundaries of model hallucinations. Crucially, to ensure reliable assessment of dynamic inputs, SAMF incorporates a structured metric suite driven by an ensemble of multi-modal oracles. Our extensive experiments reveal that state-of-the-art MLLMs exhibit significant performance degradation under fuzzing compared to conventional settings, exposing a dissociation between reasoning capabilities and factual grounding. Furthermore, we identify a helpfulness-hallucination trade-off, where reinforcement learning alignment inadvertently exacerbates sycophancy in instruction-following tasks. The framework, code and benchmark are available at https://github.com/LanceZPF/EvalHall.

Specialised retrieval agents typically surface higher quality results than general-purpose search, but selecting the optimal agent for a given query remains an open problem. Current approaches route queries based on inferred topic or intent, however intent-based selection is fundamentally limited: it does not incorporate signal from retrieved content, and cannot detect when a topically aligned agent produces low-relevance results. We address this by training a small language model via supervised fine-tuning followed by reinforcement learning to jointly perform agent selection and structured parameter generation for downstream tool calls, using a hierarchical reward function grounded in retrieval relevance along with query-agent topic alignment. This enables the model to learn task-dependent agent suitability from retrieval performance: which agents reliably yield high-relevance results for which query distributions, and when to redirect queries away from specialised agents despite surface-level topical overlap. On a targeted subset of such agent-query mismatches, the trained model achieves an NDCG@10 of 0.918 compared to 0.539 and 0.490 for two LLM baselines (Amazon Nova Lite and Claude Haiku 4.5) that route on intent alone. Overall, it achieves a mean NDCG@10 of 0.771 (+0.177 over Nova Lite, +0.219 over Haiku) with a mean selection latency of 120.1ms, an 82.4% reduction over Nova Lite.

The performance of deep learning models at scale relies heavily on how effectively high-level mathematical operations are mapped to underlying physical hardware. While high-level tensor frameworks provide flexible abstractions, their execution models inherently lack the whole-graph visibility required to maximize hardware utilization, often forcing a reliance on opaque, hand-written kernel libraries for complex operations like Attention. To bridge this gap, we present the next iteration of Nova, an automated end-to-end JIT compiler that achieves absolute control over hardware mapping by synthesizing fine-grained kernels directly from the computation's structure. In this work, we extend Nova's compilation pipeline to natively support full Transformer architectures. By capturing eager executions and unifying forward and backward passes into a single value-semantic dialect, Nova unlocks aggressive whole-graph optimizations. Rather than relying on rigid, pre-compiled library calls, Nova focuses on extensive cross-operator fusions, collapsing complex causal attention sub-graphs, element-wise operations, and memory-bound normalizations directly into single fused kernels to drastically reduce global memory roundtrips. In our evaluations training a full GPT-2 architecture on Ada 6000 GPUs, Nova demonstrates superior end-to-end throughput, averaging 441K tokens/second compared to 406K for our own eager execution and 405K for torch.compile. By drastically reducing memory-bound overheads through compiler-native fusion, Nova enables efficient full LLM compilation on modern hardware while strictly maintaining numerical parity.

Reliable and low-latency NR-V2X communications are essential for smart mobility in dense urban environments. However, limited Road-Side Unit (RSU) density, frequent non-line-of-sight conditions, and highly dynamic vehicular topologies often prevent many Connected and Automated Vehicles (CAVs) from maintaining stable single-hop connectivity. Although multi-hop relay-assisted communication can extend infrastructure coverage, selecting relay links in real time under practical flow, capacity, and connectivity constraints remains challenging. Mixed-Integer Linear Programming (MILP) yields optimal multi-hop relay decisions, but its computational complexity scales sharply with network density, limiting real-time applicability. To address this, we propose a Learning-to-Optimise (L2O) framework based on Graph Neural Networks (GNNs) for real-time NR-V2X relay selection. Vehicular communication states are modeled as attributed graphs, where CAVs and RSUs are nodes and candidate radio links are enriched with propagation-aware features. An offline MILP oracle provides optimal supervision, while an edge-aware Graph Isomorphism Network (GINE) approximates oracle decisions with near-constant inference latency. Experiments on large-scale urban datasets generated by an integrated SUMO–GEMV2 simulation pipeline show that the proposed approach achieves connectivity comparable to that of the MILP oracle while reducing execution time by orders of magnitude. The framework enables cost-effective enhancement of urban V2X connectivity by leveraging existing vehicular assets and supporting scalable, real-time NR-V2X operation in smart city environments.

Memory Manager models are pivotal in agent systems. Existing reinforcement-learning methods commonly use LLM-judged synthetic question-answer (QA) pairs: this provides useful downstream task grounding, but values memory through a sampled query distribution and a fixed reader. We propose CMI-Mem, a lightweight RL memory manager with a hybrid reward. Its extrinsic QA term measures end-task correctness, while its intrinsic Conditional Mutual Information (CMI) term evaluates the information contributed by new conversational inputs relative to the current memory state without conditioning on a sampled QA query. The two signals are complementary: QA anchors task utility, whereas CMI provides per-operation supervision for relevant, non-redundant memory construction. Experiments demonstrate improved transfer across memory-use scenarios, together with more efficient training and inference from the per-operation CMI signal. Our codes are available at: https://github.com/Wyb0627/CMIMem , and the CMI-Mem-4B model checkpoint is available at: https://www.modelscope.cn/models/wyb0627/CMIMem-4B

Effective molecular representation learning is crucial for accurate molecular property prediction. Recently, numerous self-supervised learning (SSL) approaches leveraging 3D GNNs have been developed to capture comprehensive 3D structural information for drug discovery. However, existing methods lack explicit physical constraints and are highly susceptible to geometric noise induced by coarse empirical force fields during large-scale pre-training.Furthermore, they overlook dynamic feature modulation during downstream adaptation, often resulting in catastrophic forgetting and negative transfer. To address these limitations, we introduce SenCos-GEM, a novel explicitly decoupled geometry-enhanced molecular representation learning framework that incorporates SENet-calibrated and law-of-cosines-constrained enhancements. SenCos-GEM employs a physics-guided geometric consistency loss based on the law of cosines to derive high-fidelity and mathematically invariant 3D spatial priors. In addition, lightweight Squeeze-and-Excitation (SE) modules are integrated into the backbone as task-specific adapters, while a dual-modulation prediction head combines Feature-wise Linear Modulation (FiLM) and SENet mechanisms to enable dynamic feature recalibration. SenCos-GEM demonstrates highly competitive performance across diverse classification and regression tasks on MoleculeNet benchmark, establishing new state-of-the-art results specifically on 3D conformation-sensitive regression tasks, such as FreeSolv, Lipophilicity, and QM9, achieving relative error reductions of 12.9% (RMSE), 5.3% (RMSE), and 8.2% (MAE), respectively. Moreover, our model exhibits superior capability in distinguishing stereoisomers and discriminating conformational perturbations, underscoring its robust spatial modeling performance. Collectively, SenCos-GEM represents a significant breakthrough in accurate molecular property prediction.

Micro-gesture recognition demands the detection of fleeting, spatially localized movements that are frequently overwhelmed by dominant static appearances and background noise. While Multimodal Large Language Models (MLLMs) excel at general video understanding, they inherently struggle with subtle kinematics and often rely on static posture priors. To this end, we propose GMoT, a Gated Motion-Aware Tokenization module that explicitly distills sparse kinematic evidence into a compact sequence prior to temporal modeling. GMoT dynamically spotlights action-relevant regions via spatially weighted pooling, extracts adjacent-frame temporal differencing to capture precise motion energy, and adaptively fuses these cues into the visual stream using a conservatively initialized semantic gate. To transition from simple classification to evidence-grounded reasoning, we further introduce a progressive reward-guided policy refinement paradigm, supported by a semi-supervised annotation pipeline that generates anatomically focused captions. Beyond achieving the best Top-1 accuracy among the compared methods on iMiGUE (67.32%) and SMG (73.11%), improving the Qwen3-VL-8B baseline by +6.80 and +3.11 points, our framework introduces Body-Region Grounding (BRG) Recall as an anatomical-grounding proxy conditioned on correct predictions, together with an overlapping-label cross-domain transfer protocol between iMiGUE and SMG. Extensive evaluations demonstrate that our GMoT-augmented model improves in-domain accuracy, retains clear gains under label-preserving corruptions, and improves accuracy-oriented cross-domain transfer under explicit small-split caveats while maintaining high anatomical grounding in its generated rationales.

Deep networks now subtype brain tumors on MRI about as well as specialist readers, yet accuracy is not what keeps them out of the clinic. What matters at the point of care is whether a model's confidence can be trusted to flag the cases it is likely to misclassify and defer them to a human. Deterministic estimates cannot: an auxiliary confidence head trained alongside the classifier collapses to a near-constant output that says nothing about correctness. This study proposes an uncertainty-first pipeline for four-class brain tumor MRI (glioma, meningioma, pituitary, no tumor) that reads predictive uncertainty from Monte Carlo (MC) Dropout over T = 20 passes and turns the resulting entropy into an explicit rule for deferring uncertain cases to a radiologist. We partitioned 7,200 images by perceptual-hash cluster, closing the near-duplicate leakage that inflates accuracy under naive splitting, and evaluated the pipeline on ViT-B/16 and ResNet-50 across five seeds along three axes: discrimination, calibration, and selective prediction. Both discriminate strongly (macro-AUC 0.994; accuracy 0.962 and 0.964), and no seed separates them (0 of 5 significant, p < 0.05), so the result is driven by the uncertainty pipeline, not the network. A single temperature scalar pulls the deterministic softmax into tight calibration (expected calibration error 0.016-0.020), and deferring the most uncertain 5% of cases lifts accuracy on the rest to about 0.98 on both (area under the risk-coverage curve 0.010-0.011). MC-Dropout uncertainty here is thus calibrated, non-collapsing, and directly actionable through a concrete deferral rule, providing an architecture-agnostic basis for calibrated, defer-to-human brain tumor MRI triage under internal validation.

Healthcare spans high-stakes communication, expert reasoning, and workflow execution, yet specialized language models that cover these use cases together remain limited. A healthcare model must handle patient consultation, clinical reasoning over text and images, interactive diagnosis, and electronic health record (EHR) tool use. These capabilities fail in different ways, and a narrow update for one task can degrade another. We present Cura 1T, a healthcare foundation model trained through recursive self-improvement (RSI). In each RSI round, the RSI harness runs the current model on healthcare benchmarks, evaluates the trajectories to locate capability gaps, and refines the training mixture by synthesizing training data. On 6 healthcare benchmarks, Cura 1T scores highest on MedAgentBench, HealthBench Professional, HealthBench Hard, MedXpertQA text, and AgentClinic, and second on MedXpertQA multimodal. It preserves performances on out-of-domain reasoning and agentic benchmarks including AIME, GPQA-Diamond, and $\tau^2$-Bench.

We report a way to make a frozen small language model both more capable and dramatically cheaper at once, without changing any weights. Verified knowledge is deposited once as a byte-exact key-value (KV) state artifact and later restored, by graft, into a fresh inference context. The restore is bit-exact: under a pinned deterministic configuration, the grafted logits are byte-for-byte identical to a fresh computation (SHA-256 equality), with zero KL divergence and 100% argmax agreement over fifty samples. We show that own-position graft is the unique numerically exact operating point on a model with floating-point rotary encoding, and we verify byte-exactness on two model scales (12B, 31B) and two GPU targets, one through a pre-registered replay. On AIME 2025, a frozen Gemma-4-12B moves from 80.0% to 93.3% once a verified solution library is grafted, above its own 77.5% and its 31B sibling's 89.2% published anchors. On the recurring case, eight problems the base model never solves within a 401,026-token budget are answered from cached verified solutions in 61 total decode tokens, a factor of 6,574 fewer tokens and about 8,700x less energy; the capability claim proper rests on held-out transfer (7 of 7 at 31B). The same byte-exact store widens usable context from 32,768 to 2,854,766 tokens at zero extra accelerator memory, and moves byte-identical between machines of the same architecture. We describe the system at the behavior level; the engine is proprietary, and every reported number is backed by committed input and output hashes so the scoring can be re-checked without it.

A depth-recurrent transformer applies a weight-tied core a variable number of times, and prior work has shown that training with a randomized recursion count yields one checkpoint usable across a range of inference depths. We ask what such a model actually computes per token, and measure it directly. On a 135M-class model trained on FineWeb-Edu, the recurrent state converges to a per-token fixed point: mean successive-output KL divergence falls from 3.9e-1 at the second loop to 8.5e-6 by the sixteenth, and per-token state change decays in step. Crucially, this convergence is not uniform across tokens. The median token converges by loop six, while approximately 10 percent of tokens continue to update at the training-mean depth of eight, and mean convergence depth is ordered by token type (whitespace shallowest, content words deepest). This per-token variation is the central object of the paper. We show it is directly readable and that reading it outperforms learning to predict it: a training-free rule that halts each token once its output stabilizes attains uniform depth-8 quality at 4.94 average loops (a 38 percent reduction in average depth) and matches uniform depth across the average-depth range, whereas a linear router trained on convergence labels harvested from the same model requires nearly full depth and yields no reduction. The elasticity that makes this possible reproduces here as background (validation loss decreases monotonically from 3.80 at one loop to 3.20 at eight and remains stable to 32 loops). We report average depth as a FLOP proxy with a three-point wall-clock bracket rather than a realized speedup, make no FLOP-matched parity claim, and note that the allocation results are established at a single scale and seed. The complete study runs on a single RTX 4090 in approximately 100 GPU-hours.

In recent years Flow Matching has become a prominent method for generative modeling robot motion generation. In its generic form Flow Matching is an ODE-based neural sampler that is trained by regressing empirical flow fields associated with motion samples as data. However, in robot motion generation we often have additional constraints that might not be present in the collected data. The majority of current approaches train the flow on the available data and use inference-time guidance to enforce task-specific constraints. To address this mismatch, we propose ConFlow, a constraint-guided flow matching framework that incorporates constraint information directly into the training objective via differentiable barrier or cost functions. To address design specifications such as smoothness and boundary conditions, we propose replacing the standard Gaussian source distribution used in flow matching training with a conditional Gaussian Process. Our approach also uses infeasible demonstrations as negative supervision, improving constraint satisfaction without requiring additional expert data. Experiments on a two-robot navigation task demonstrate that ConFlow achieves lower collision rates and higher trajectory quality than standard flow matching baselines, with or without inference-time guidance. These results validate training-time constraint integration as an effective approach to closing the training–inference gap in generative motion models.