Decision trees are prized for their interpretability and strong performance on tabular data, but popular greedy top-down induction algorithms can yield suboptimal and unnecessarily complex structures. Optimal decision tree methods address this through global optimization, yet remain restricted to axis-aligned threshold splits, which limit the expressivity of each node and often force deep, complex trees to capture non-linear feature effects. Shape Generalized Trees (SGTs) generalize threshold splits to learnable univariate shape functions, improving expressivity and enabling more compact trees. However, existing SGT induction algorithms are greedy and offer no optimality guarantees. In this work, we introduce Literati, the first algorithm for optimal SGT induction. We propose a novel AND/OR graph formulation of the problem that jointly optimizes tree structure and shape function complexity. To solve this AND/OR graph, we develop an AO*-based algorithm with two enhancements that improve anytime performance while preserving optimality: a secondary heuristic for OR-node selection and a round-robin policy for AND-node exploration. Across 24 real-world datasets, Literati achieves higher training and test accuracy than state-of-the-art tree approaches.
Multi-step LLM reasoning lacks a machine-recheckable ledger: discarded reasoning paths leave no auditable record. We propose the Deposon scattering layer, which binds each node of an LLM-generated concept-decomposition graph to a two-parameter Deposon state; paths undergo three-channel scattering -- transmission, reflection, irreversible dissipation -- obeying T+R+A=1 for arbitrary parameters, with a maximum per-path energy-audit deviation of 2.2E-16 (machine epsilon). We report all three evidence tiers honestly. On synthetic trap benchmarks the path-filtering gain is closed (pre-registered): unified reaches 100% versus a decoy-capture baseline at 7%/10%. On real benchmarks the layer is indistinguishable from a trivial six-keyword rule filter (GSM8K 0.87 >= 0.85, McNemar p=0.5; StrategyQA 0.899 = 0.899); no difference is detected here, so we sharpen the claim to "the differential value lies solely in machine verifiability." Fusion yields a second negative result: convex combinations with a semantic prior never improve (physics 0.484 -> 0.452), and the apparent lambda=2 gain is an anti-field artifact; any fusion gain must be nonlinear. Modeling the reverse dynamics as a potential game on the graph, we evidence an auditable scalar's monotonicity and near-gradientness and quantify the empirical coordination ratio (ECR). The three formalized dynamical-equivalence propositions (P1a/P1b/T-P1c) are falsified under the pre-registered kill protocol, and the potential-game claim is downgraded to approximate (cyclic-graph median residual 0.669): only consistency-level evidence survives at the dynamical level. Code: github.com/zeroandcat/Deposon.
Fair clustering aims to prevent algorithmic bias against protected groups. We study Euclidean fair $k$-means, in which the proportion of each protected group in every cluster lies between prescribed lower and upper bounds. Because these constraints violate the nearest-center assignment property, center selection is coupled with point assignment. We introduce the drift-regularized relaxation, a new convex relaxation solved by a modified Frank-Wolfe method with linear-programming iterations, whose optimality condition controls the cost of converting its solution into $k$ fair clusters. Combined with any $\rho$-approximation algorithm for weighted $k$-means, the relaxation yields an exactly fair fractional solution of cost at most $2\rho+1+\epsilon$ times the fractional optimum, improving the previous ratio $1+4\rho+\epsilon$. In fixed dimension, this improves the best known ratio from $5+O(\epsilon)$ to $3+O(\epsilon)$. The fractional solution can be rounded to an integral one without changing the centers or increasing the cost, at a bounded additive violation of the fairness bounds. We also show that a $(3-\delta)$-approximation is NP-hard for every fixed $\delta>0$, even on the line with fractional assignments, so $3$ is the exact approximation threshold for fractional fair $k$-means in fixed dimension. The same guarantees hold for $k$-sparse Wasserstein barycenters, a special case of fractional fair $k$-means. More generally, they hold whenever every cluster must satisfy given homogeneous linear inequalities in the fractions of its points, such as a bound on the ratio between two groups or on the average of a numerical attribute in every cluster.
Model pruning is widely used to compress deep neural networks, reducing memory and computational requirements with minimal impact on aggregate performance. However, its effect on model behavior remains poorly understood, particularly for long-tailed medical datasets where rare but clinically important conditions are underrepresented. Furthermore, it remains unclear whether pruned models preserve reliable explanations of their predictions. To address this gap, we present a systematic study of long-tail forgetting and explanation reliability under model pruning. Across two long-tailed medical imaging datasets, two CNN architectures, four pruning methods, and sparsity levels up to 95\%, we evaluate predictive performance, explanation stability, and explanation faithfulness. Our results show that predictive performance exhibits a strong frequency-dependent trend, with lower-frequency classes generally experiencing earlier and larger degradation than higher-frequency classes. In contrast, explanation stability and faithfulness are influenced primarily by the pruning strategy, with gradient-informed methods preserving explanation reliability more effectively under aggressive compression. Qualitative and mechanistic analyses further indicate that explanation degradation is primarily associated with the collapse of class-discriminative gradients rather than the disappearance of feature activations. These findings suggest that model compression should be evaluated beyond aggregate performance. Incorporating class-aware and explanation-aware evaluation reveals failure modes that would otherwise remain hidden, while moderate sparsity levels provide a practical balance between compression, predictive performance, and explanation reliability.
Learning on dynamic graphs is difficult when changes in the underlying network are only partially observed. Acquiring current graph information incurs observation and computational costs, making complete updates impractical under limited resources. This paper focuses on budgeted task-aware acquisition on dynamic networks, where a model needs to decide which stale graph information to refresh for a downstream task. We propose Scout, a lightweight framework that learns the task value of querying each node from the maintained graph and observation history. Our evaluation covers one synthetic and 4 real-world dynamic networks, 2 downstream tasks, 9 acquisition baselines, and several query budgets. Scout achieves the highest mean downstream performance in 15 of the 21 benchmark-budget settings and ranks within the top two in 19 of them. Its average rank is 1.4, compared with 3.2 for the strongest baseline. On the same dynamic network, task-matched acquisition improves link-prediction AUC by 0.005-0.007 and node-classification accuracy by 0.007-0.017 over task-mismatched acquisition. These results show that useful graph observations depend on the downstream task and that limited observation budgets can be allocated effectively by learning directly from downstream utility.
Eye-tracking data are expensive to collect, requiring specialized hardware and controlled laboratory conditions, and difficult to share due to privacy regulations, which together limit the quantity and diversity of gaze data available for research. We investigate unconditional gaze generation using diffusion models in position and component-wise velocity representations. The models are trained and evaluated on eight-second visual-search trajectories from 28 participants, using a split of 23 training participants and five held-out validation participants, with three training seeds. Evaluation includes kinematic distributions, trajectory geometry, event statistics, and a temporal Wasserstein GAN with gradient penalty as a learned comparator. Across eight position-derived features, mean Jensen-Shannon divergence is $0.016 \pm 0.004$ for the position model and $0.037 \pm 0.002$ for the velocity model, with standard deviations calculated across training seeds. These scores do not establish that representation choice caused the difference. A separate comparison using the same evaluation method shows that the ordering of the models varies across training seeds. The position model also produces smaller feature discrepancies than a kinematic Markov baseline. Differences in saccade counts between real and generated data depend strongly on the velocity threshold used to detect events. The results provide quantitative reference measurements for unconditional gaze generation on this visual-search task and show how evaluation choices affect conclusions about generated gaze.
Multimodal Earth Observation models often fuse radar and optical inputs early, limiting sensor-conditioned processing, or maintain separate sensor streams through the encoder, increasing computational cost. We present MEOX (Multimodal Earth Observation with eXperts), which delays fusion until after a shared mixture-of-experts block that processes Sentinel-2 and ascending and descending Sentinel-1 streams independently. Co-located patches are then fused with validity-aware weights into a single sequence processed by fourteen shared blocks, reducing the sequence length from 3N + 5 to N + 5 tokens for N patches. We pretrain the 2.939M-parameter encoder on 1.228M MMEarth64 samples using masked radar-optical reconstruction. Under the reported frozen-transfer protocols, MEOX reaches 90.56% average class accuracy on EuroSAT and 64.42% mean intersection over union (mIoU) on cashew segmentation, 2.26 and 5.02 points above the strongest reported cross-sensor mixture-of-experts (CSMoE) variants. Full-encoder adaptation increases BigEarthNet micro-average precision from 57.22% to 72.95%. In a controlled same-corpus test using synthetic random patch removal, masking 75% of Sentinel-2 patches reduces WorldCover mIoU from 30.45% to 26.53% when both Sentinel-1 streams remain available; with Sentinel-2 alone, mIoU falls to 13.81%. For identical zeroed optical patches, declaring them missing through the validity mechanism improves mIoU from 19.42% to 26.53%. These results characterize a compact route-then-fuse encoder that retains sensor-conditioned processing and uses complementary radar observations when optical inputs are incomplete.
Designing and authoring high-performance custom kernels for accelerators is a complex task that requires deep hardware-level expertise. Large Language Models (LLM) can be leveraged together with real-time compiler feedback to build agentic systems for kernel generation. In this work, we present MaxKernel, a multi-agent system that implements three distinct paradigms for TPU kernel development: (1) a Human-in-the-Loop (HITL) agent for collaborative, step-by-step design; (2) an Autonomous (Auto) agent that executes a fully automated, metric/trace-driven optimization loop; and (3) a Graph-Based Autonomous Search that scales the Auto agent for global exploration of the design space. All three paradigms leverage a shared pool of specialized sub-agents to handle planning, implementation, self-debugging, testing, and hardware profiling. We evaluate MaxKernel on JaxBench, a comprehensive suite of 50 diverse kernel tasks for TPUs, alongside complex, real-world workloads from state-of-the-art open-source models. We demonstrate that MaxKernel consistently generates highly optimized implementations, matching expert hand-tuned baselines and delivering significant performance across the benchmark. Our agent is open-sourced and available https://github.com/AI-Hypercomputer/accelerator-agents/tree/main/MaxKernel.
Modern misinformation is often heard before it is read, yet fact-checking systems are still evaluated mainly on clean written claims. Spoken dialogue remains different even when systems operate on transcripts: claims may be distributed across speakers and turns, depend on prior context, and become harder to verify when Automatic Speech Recognition (ASR) errors distort the available text. Prior spoken dialogue fact-checking resources are small, English-centric, or focused on annotation rather than end-to-end benchmarking, leaving no large multilingual benchmark with paired speech and turn-level labels. We introduce TRILOGUE (TRIlingual spoken diaLOGUE fact-checking), a large-scale trilingual benchmark of source-grounded spoken dialogues in English, Russian, and Kazakh. It contains nearly 12K dialogues, 187K turns, and 390 hours of paired audio with ASR transcripts and word-level timestamp alignments across all three languages, including nearly 5K human-recorded Russian and Kazakh dialogue files. TRILOGUE supports claim check-worthiness detection, source-article evidence retrieval, and claim verification with claim-only, gold-evidence, and retrieved-evidence inputs. Baselines show that ASR degradation and cross-lingual transfer remain challenging, especially for Kazakh, while retrieved source evidence substantially narrows the gap to gold-evidence verification.
Longitudinal prediction from electronic health records (EHRs) is limited by the sparsity and irregularity in patient trajectories, and knowledge augmentation with external knowledge graphs (KGs) offers a promising way to alleviate these issues. However, KG expansion often follows predefined rules, with limited control over whether and how to augment each visit as a patient's state evolves. We propose ReTA, a Reinforcement learning-based dynamic Topology Augmentation framework that casts KG import as a per-visit, budget-aware policy. ReTA first constructs an offline refined pool of KG-grounded templates, then learns a policy to select one augment action per visit from three options: Soft Import, which enriches node features without modifying graph topology, Hard Import, which grafts a compact KG subgraph onto the visit graph to create message-passing shortcuts, and Skip, which leaves the visit unaugmented when the base encoder is already confident. To stabilize learning, ReTA employs a decoupled encoder that processes semantic and structural signals in separate channels and fuses them via adaptive gating. Experiments on MIMIC-III and MIMIC-IV across diagnosis prediction, mortality, and readmission show that ReTA consistently outperforms strong baselines while remaining efficient, with larger gains under sparse supervision. ReTA transfers across datasets and knowledge graphs and yields interpretable augmentation patterns.
Automatic Target Recognition (ATR) in Synthetic Aperture Sonar (SAS) is often underpinned by deep neural networks (DNNs). Recent research has focused on relatively low-capacity DNNs (i.e., relatively fewer trainable parameters) to avoid overfitting on the small training datasets that are typically available for SAS. However, recent research has found that increasing DNN capacity often improves performance, even when training data is scarce. Motivated by this emerging finding - termed the double descent phenomenon - this work investigates the use of especially high-capacity DNNs for SAS-ATR. We evaluate the performance of several state-of-the-art DNN architectures as we increase their model capacity. We examine performance in a realistic leave-one-geo-location-out cross-validation scheme. We find that increasing model capacity often increases performance substantially, and is never harmful. All of our large models significantly outperform two baseline models from recent publications. Training high-capacity models on small dataset however requires careful regularization, and we also perform systematic experiments showing the impact of several regularization strategies: pretraining on other imaging modalities, data augmentation (including some new to SAS), DropPath, and weight decay. We show that each of these strategies independently improves the performance of the large models.
Multi-model LLM serving uses routers, cascades and agent workflows to move requests across models with different costs and capabilities. Existing serving systems reuse KV caches only within one model, so every switch forces the target to re-prefill history the source already processed, adding time to first token (TTFT) that grows with history length and target size. Cross-model KV transfer can replace this prefill, yet prior methods are slower than re-prefill end to end. We present CacheBridge, the first serving system to accelerate heterogeneous multi-model workloads through cross-model KV transfer. It co-designs a compact, attention-calibrated KV mapper with systematic optimization for online multi-model serving, which makes model switch efficient on latency and memory cost. In Qwen3 14B$\rightarrow$32B agent serving on $\tau$-bench, it reduces mean TTFT by $30$ to $98\%$ versus re-prefill at $1$ to $24$ QPS with comparable task success. Across five model pairs from three families, it uses $8$ to $12\times$ less mapper memory than Closed-Form Mapper and retains on average $97.94\%$ and $102.94\%$ of re-prefill task quality on chat/reasoning and RAG workloads, respectively. The promising results have opened new opportunities for efficient multi-model LLM serving through cross-model KV reuse.
Recovering explicit symbolic solutions of partial differential equations (PDEs) from physics-informed neural networks (PINNs) is attractive, yet selecting and fitting expressions against neural predictions can lock the recovered solution into the teacher's approximation error. We propose DeSyR (Decoupled Symbolic Recovery) to break this dependence. DeSyR uses PINNs only to discover candidate structures and initial coefficient estimates; once a structure is fixed, the governing equations and constraints alone determine the refined coefficients. Multiple candidates with different fitting-error--complexity trade-offs are retained, evaluated after physics-only refinement, and reduced to the simplest expression within a physically near-optimal set. Joint refinement extends the same principle to coupled fields. For fixed candidate parameterizations, affine error identities and local nonlinear analysis show how teacher errors enter fitted coefficients and establish conditions under which physical residuals control coefficient and solution errors. Across seven benchmarks covering scalar evolution equations, multidimensional elliptic problems and coupled incompressible flows, DeSyR attains mean relative solution errors on the order of $10^{-14}$ or smaller and lower reported mean physical residuals than PINNs, physics-informed symbolic networks and teacher-fitted symbolic distillation. With the expression structure held fixed, estimating coefficients from physical constraints further yields more accurate frequency estimates and solutions than fitting teacher predictions. DeSyR therefore turns neural approximations into guides for structure discovery, while physical constraints determine the final explicit solution.
A direct test of an LLM judge of agent trajectories injects faults into correct runs and reports recall, per fault type or by whether the fault broke the environment outcome (loud) or not (silent). Such recall can credit a judge with detection it does not have; paired discrimination, its flag rate on the faults minus its rate on the clean runs they came from, exposes this. Our testbed, a deterministic support desk with a scripted oracle and a one-step fault injector, labels all 400 trajectories exactly. A 14B judge shown only the request and final reply scores 34% to 76% recall on four fault types that leave the reply unchanged. There its input is the clean run's, so its paired discrimination is zero and that recall is its flag rate on clean runs. Splitting by outcome survival does not fix this: its loud recall of 84% is a paired +0.393 and its silent recall of 45% a paired +0.048, all from the two fault types that change the reply. Told to check each step, the same model flags every fault of those four types and 0 of 100 clean runs (95% CI up to 3.6%). It does not reliably check the reply: of four invented promises it flags one every time and the other three once in 42 faults. Shown every step but asked only about the reply, it still reaches a paired +0.69 on reply-unchanged faults, against +1.00 when told to check each step. We recommend reporting paired discrimination against clean parents, split by whether the fault reaches the judge's input and by outcome survival, and release the testbed, raw verdicts and analysis pipeline.
Koopman autoencoders (KAEs) seek a higher-dimensional latent representation in which nonlinear dynamics evolve linearly. However, many interesting systems have multiple basins of attraction, and both theoretical and empirical work has shown these multibasin systems cannot generally admit a single finite-dimensional global Koopman embedding under standard assumptions. We posit that encoders with a sparsity-inducing objective encouraging few active latent coefficients will provide latent supports as an inspectable basin-modeling principle for Koopman autoencoders. We use these encoders producing sparse latents in training Sparse Koopman Autoencoders (SKAEs) without basin labels or other regime annotations, and treat the learned latent supports as model-produced regime variables after training. Across a range of procedurally generated multibasin systems and chaotic flows, we show that SKAEs have superior forecasting performance compared to dense-latent KAEs. We also perform a mechanistic study that shows latent supports produced by SKAEs are both essential for the quality of the representation and useful for identifying basins on basin-interior states, whereas dense-latent KAEs collapse to an uninformative single support family. These results identify sparse latents and their corresponding supports as label-free, interpretable regime variables for Koopman learning in nonlinear systems with multiple local dynamical laws.
Diabetic retinopathy (DR) is a leading cause of preventable blindness, and automated screening must be both accurate and trustworthy. We study an explainable DR classification framework built on vision foundation models and three transfer learning strategies. Three backbones, DINOv2, CLIP, and Vision Transformer (ViT), were adapted by full fine-tuning, linear probing, and Low-Rank Adaptation (LoRA). Models were trained and internally evaluated on the ODIR dataset and externally evaluated on APTOS to assess generalization. DINOv2-LoRA achieved the highest internal AUROC (0.758), whereas DINOv2 and ViT full fine-tuning achieved the highest external AUROC (0.920). Probability calibration was assessed with reliability analysis after isotonic regression. For explainability, Grad-CAM and HiResCAM maps were compared with expert-annotated lesion masks from the IDRiD dataset using Dice, Intersection over Union (IoU), and Pointing Game metrics. The results show that foundation models, particularly DINOv2, provide strong predictive performance, and that LoRA is a parameter-efficient alternative to full fine-tuning. Spatial overlap between explanation maps and annotated lesions was limited (mean Dice of about 0.13 for the combined lesion mask), which highlights the value of quantitatively testing whether model attention corresponds to clinically relevant retinal lesions.
While text-based hallucination detection is well studied, reference-free detection of factual alterations in speech remains underexplored, especially for low-resource languages. Our spoken benchmark comprises 12,013 English, Russian, and Kazakh news samples with three synthetic alteration types and three severity levels, pairing source articles with rewrites as text, synthesized audio, and ASR transcripts. We add 290 fact-checked misinformation items collected in Russian (225) and Kazakh (65), translated into the other language and rendered through the same TTS-ASR pipeline. We evaluate fine-tuned multilingual encoders and zero-shot multimodal decoders on text, transcripts, and audio. Detectors receive only target inputs without source articles or external evidence; the task evaluates reference-free classification rather than evidence-grounded verification. Encoder degradation from source text to transcripts generally tracks per-language ASR error on the binary task. Among decoders, only Gemma-3n exceeds the binary majority-class baseline in macro-F1, on transcripts only; the other four fall below their respective baselines. Comparisons between audio and transcripts are confounded by differences in evaluation coverage and class balance. Synthetic-trained detectors achieve 0.82--0.88 macro-F1 on real-world misinformation source text; Russian provenance analysis reveals veracity-related and model-dependent machine-style signals, a key confound in synthetic hallucination benchmarks.
Vision-Language Models (VLMs) based GUI agents stand to benefit significantly from online reinforcement learning (RL). However, their training is bottlenecked by two fundamental issues: current data synthesis methods for GUI Agents rely on specific environments and struggle to generate diverse data, while existing evaluators either suffer from limited scalability or provide inaccurate and unreliable reward signals. To overcome these challenges, we introduce GSAR (Goal-State-Anchor Reward), a RL reward framework that supports scalable task generation and delivers reliable reward signals for stable and efficient policy optimization. Our approach features self-evolving data synthesis, which produces multiple environments through task execution and generates diverse tasks and goal states. Complementing this, a state-anchor mechanism automatically annotates task-relevant UI elements in successful goal states as reference anchors. During RL training, these reference anchors provide accurate, scalable reward signals that substantially enhance efficiency. Extensive evaluations demonstrate that our framework achieves over 90% accuracy on offline trajectory verification and performs closest to rule-based methods. Furthermore, agents trained using our reward framework exhibit strong performance on both AndroidWorld and our constructed benchmark, establishing a scalable approach for GUI agent training.
Personal AI agents consume external content such as web pages and SNS feeds, and retain selected information in persistent memory. As a result, external content has become a new attack vector for memory poisoning, which steers an agent's later actions toward an adversary's bias. However, most existing attacks embed a malicious instruction in external content to make the agent retain the poisoned information, and prompt-injection filters can detect such content. We observe that agents access external content primarily through skills while curating a subset of the content into their memory. Leveraging this curation path, we present SBIA, a Stealthy Bias Injection Attack that introduces adversarial bias through ordinary external content while relying on the agent itself to perform the memory write. The adversary posts crafted comments under hot SNS feeds and distributes an SNS feed summarization skill whose curation rule admits these comments into memory. Comment cloaking expresses the bias as an ordinary opinion rather than an instruction, comment watermarking allows the skill to identify the crafted comments, and topic anchoring records them so that later related requests recall them. On BiasBench with 7,000 crafted SNS comments and seven LLMs on OpenClaw, SBIA achieves attack success rates of at least 96% across models and even 91% on GPT-5.5 in four downstream tasks. Despite this effectiveness, its skill passes all eight evaluated malicious-skill scanners and ClawHub's official skill scanning pipeline. We further propose a memory boundary defense that detects 82.1% of injected entries at a 5.6% FPR. Our findings show that external Web content and third-party skill marketplaces form a new attack surface for personal AI agent memory.
We propose VT-MUSE, a Multimodal Unified SEquential representation learning framework for visuotactilemanipulation. Existing approaches often encode visual and tactile observations independently before fusion, limiting their ability to capture fine-grained cross-modal dependencies. Moreover, most methods focus on observations at the current time step and overlook the temporal evolution of contact. VT-MUSE addresses both limitations through a two-stage representation learning framework. In Stage I, modality specific encoders are jointly adapted via cross-modal temporal alignment and masked-view consistency. In Stage II, a conditional variational latent model processes masked visual sequences together with full tactile histories. Auxiliary decoders reconstruct the masked recent visual observations and predict tactile depth changes, encouraging the latent representation to retain both global visual context and local contact dynamics. The learned representation is subsequently integrated into a lightweight Transformer policy through gated cross-attention. On the simulation benchmark, VT-MUSE outperforms the strongest baseline evaluated on all tasks by 11 percentage points and also achieves substantial improvements in real-world experiments.