On-Policy Distillation (OPD) is a fundamental technique for efficient post-training of large language models (LLMs), with broad applications in agent learning, multi-task enhancement, and model compression. However, OPD training becomes unstable when the teacher and student distributions differ substantially, as teacher supervision on student-generated tokens may yield unreliable policy gradients and even cause optimization failure. This work addresses reliable on-policy token-level supervision through credit assignment strategies, and proposes Trust Region On-Policy Distillation, TrOPD. It features the following characteristics: 1) Trust-Region On-Policy Learning: TrOPD performs OPD only in regions where the teacher provides reliable supervision, mitigating the optimization difficulty of the K1 reverse-KL estimator under distribution mismatch. 2) Outlier Estimation: For outlier regions, we explore gradient clipping, masking, and forward-KL estimation to reduce the adverse effects of unreliable supervision. 3) Off-Policy Guidance: The student continues generation from teacher prefixes and uses forward KL to imitate off-policy guidance, encouraging on-policy exploration toward reliable regions. Experiments show that TrOPD consistently outperforms SoTA OPD baselines, including OPD, EOPD, and REOPOLD, across mathematical reasoning, code generation, and general-domain benchmarks.

Retrieval-Augmented Generation (RAG) systems are widely deployed and increasingly influential, but their reliance on external corpora exposes new security risks from poisoned retrieval content. Existing RAG attacks are largely focusing on individual queries or narrow topic-local query sets, which limits their practical reach and offers limited camouflage in real-world settings. In this paper, we introduce discourse-level opinion manipulation, a new threat model in which coordinated influence across a semantic query network induces opinion shifts over a holistic, multi-topic query space. We formalize this threat in a black-box setting and propose DiscourseFlip, an agentic, graph-guided attack that dynamically allocates a limited poisoning budget to maximize discourse-level opinion deviation. Extensive experiments demonstrate that DiscourseFlip consistently induces targeted opinion shifts across the contextualized query network and significantly outperforms existing baselines in terms of coverage and effectiveness. User studies further confirm that DiscourseFlip is effective while remaining well camouflaged from user detection. Moreover, systematic analyses show that existing mitigation strategies are ineffective against discourse-level manipulation, underscoring the urgent need for more robust and adaptive defenses to address discourse-level vulnerabilities.

Agent-memory frameworks -- mem0, Letta/MemGPT, Cognee, Zep/Graphiti, MemoryOS, MemTensor -- each ship their own SDK, storage layout, and operational vocabulary. There is no shared wire format: every integration is bespoke, every migration rebuilds memory from scratch, and no framework ships a governance surface that lets a human review writes before they enter long-term storage. We present memorywire, a JSON-Schema 2020-12 wire format for five memory operations (remember, recall, forget, merge, expire) over four memory types (semantic, episodic, procedural, emotional), with a MemoryStore interface, a fan-out router, and an optional HITL governance channel. We describe an open-source reference implementation with five backend adapters (sqlite-vec, mem0, Letta, Cognee, pgvector); a microbenchmark on a 100-fact / 50-query labelled corpus (42 with non-empty gold ids + 8 no-match probes) achieving recall@5 = 1.000 on the 42 gold-id queries with ingest p50 = 37.8 ms and recall p50 = 40.6 ms; an adversarial-fusion experiment showing Reciprocal Rank Fusion holds recall@5 = 1.000 across a 1-of-N rank-0 injection sweep (K in {0, 5, ..., 50}) where max fusion collapses to 0.500 with 80% leak at K >= 5; and a 16-scenario cross-adapter conformance suite passing 68 of 80 cells with zero failures. The contribution is not a new algorithm; it is a packaging of established components (RRF, FSMs, STM/LTM consolidation, diff-and-approve workflows) into a venue-neutral protocol with an empirically validated reference, positioned to compose with the Model Context Protocol rather than compete with it.

Visual inspection remains the dominant quality-control practice in woven and tufted carpet production, yet it is slow, subjective, and inconsistent at the line speeds and widths of modern looms. We present a design proposal for an in-line machine-vision system whose primary purpose is twofold: to inspect the carpet web in real time and, equally importantly, to systematically collect and label images of defect patterns so that increasingly capable quality-control models can be trained over the life of the installation. The proposal is grounded in a concrete industrial setting: a Six Sigma (DMAIC) project at a woven-carpet production facility that anticipated a production bottleneck following the installation of additional weaving machines, with a substantial baseline defect rate and significant financial exposure associated with quality failures. We describe an imaging subsystem based on synchronized line-scan cameras with combined bright-field and grazing illumination, derive the resolution and throughput requirements needed to resolve fine structural defects across a multi-metre web, and define a carpet-specific defect taxonomy. We then lay out a staged modelling strategy that begins with unsupervised anomaly detection trained on defect-free material, following the paradigm exemplified by the carpet category of the MVTec Anomaly Detection benchmark, and matures through a human-in-the-loop annotation flywheel into supervised detection and segmentation models. Finally, we connect detection performance to the DMAIC objectives, showing how reductions in escaped defects translate into improved process quality and process sigma levels. The contribution is an end-to-end, deployable blueprint that treats data collection as a first-class engineering objective rather than an afterthought.

For low-altitude Unmanned Aerial Vehicle (UAV) autonomy, 3D spatial understanding is not merely a perception objective, but the safety interface between human instructions and physical flight. In human-scale urban airspace below 20 meters, thin geometry, occlusions, vegetation, and urban clutter define whether an aerial agent can safely enter the space ahead. However, existing UAV datasets mainly provide 2D annotations or 3D boxes, while driving-oriented occupancy benchmarks assume stable ground-level sensor rigs. Both miss the defining regime of low-altitude flight: a front-facing monocular camera observing occupied and free space from a moving aerial body with frame-wise changing 6-DoF pose and camera extrinsics. To bridge this gap, we introduce SkyShield, to the best of our knowledge the first front-view monocular semantic occupancy benchmark for urban UAV flight below 20 meters. Built on CARLA, SkyShield contains 36K front-view UAV samples across diverse urban scenes and weather conditions, pairing each image with frame-wise 6-DoF UAV pose, frame-wise dynamic camera geometry, UAV states, and front-frustum semantic occupancy labels. We further propose KAR-mIoU, a UAV-centric and dynamics-aware metric that re-weights voxel-level evaluation by kinematic reachability and time-to-collision, revealing safety-critical risks hidden by conventional mIoU. To tackle this challenging new setting, we provide SkyOcc, a geometry-first monocular baseline that integrates frame-wise UAV attitude into projection, fuses temporal occupancy features, and applies safety-prior optimization to preserve sparse collision-critical structures. Together, SkyShield, KAR-mIoU, and SkyOcc establish occupancy as a safety interface for low-altitude aerial autonomy. Code and dataset will be released publicly.

Learning long-range non-stationary temporal patterns remains a core challenge for modern sequence models, particularly in strict streaming settings. In these settings, data arrive sequentially and must be processed in a single pass without simultaneously revisiting past observations. Standard architectures, including recurrent neural networks and transformers, are constrained by either truncated backpropagation through time horizon or explicit input window length for long range credit assignment. To address these limitations, we propose SHARP (Sleep-based Hierarchical Accelerated Replay), a framework that decomposes temporal learning into two complementary components: a memory module that accumulates a structured history of past inputs, and a pattern-recognition module that operates over this memory. This separation enables resource- and compute-efficient adaptation to non-stationary dynamics by eliminating the need for backpropagation through time across many steps for long-range credit assignment. Inspired by the accelerated replay observed in rodents during slow-wave sleep, SHARP incorporates offline (sleep) phases in which temporally structured memory traces are replayed in an accelerated form and integrated into higher-level memory representations, improving long-range context retention. Through controlled simulations and ablation studies, we characterize the key properties of the proposed framework. In benchmark datasets such as text8 and PG-19, we demonstrate that SHARP improves over recurrent baselines by retaining next-token predictive performance on previously seen data while continuing to learn from the current stream and generalizing to future unseen data. These gains are enabled by its hierarchical structure, which yields an exponentially increasing effective temporal context with only linear-time computational cost.

Sparse autoencoder (SAE) features are increasingly used to interpret language models, with auto-generated natural-language labels serving as the primary interface for understanding what each feature represents. We ask whether these labels generalize: does a feature labeled for a concept actually track that concept across languages and scripts? Using Serbian digraphia as a controlled testbed--the same language written in both Latin and Cyrillic via deterministic transliteration--we first find that SAE feature sets activated by the same content in different languages, scripts, and wordings share substantial overlap (mean Jaccard 0.39 vs. 0.13 random baseline, peaking at 0.57), suggesting genuine cross-lingual semantic features. We then test whether auto-interpretation labels keep pace. They often do not: features whose labels describe semantic content miss the same meaning in Serbian up to 4x more often thanwithin English, and miss Serbian Cyrillic more than Serbian Latin--two scripts that are deterministic transliterations of each other--suggesting the failures align with how well each form is represented in training. The gap grows with network depth, yet the labels give no indication that they fail. These results suggest that auto-interpretation labels may reflect a feature's behavior on well-represented inputs rather than the concept itself.

Human Activity Recognition (HAR) from ambient sensors enables smart-home applications such as health monitoring and assisted living. In realistic deployments, however, sensor events arrive as a continuous stream and activity boundaries are unknown. Sliding-window inference therefore produces many windows that straddle transitions and contain mixed activities, creating boundary contamination that violates the pre-segmented instance assumption used by most benchmarks and models. Moreover, many pipelines under-use spatial context by treating sensor IDs as independent tokens. We present LastAct, a trajectory-centric framework for streaming smart-home HAR that targets the most recent activity under mixed windows while explicitly modeling spatial structure. LastAct projects sensor events onto the home floorplan to form a layout-aligned trajectory image sequence that preserves spatial continuity. A lightweight gate identifies contaminated windows, and a boundary localizer estimates the last transition to enable boundary-guided masking that emphasizes post-boundary evidence and suppresses stale context. For efficiency, we reuse a precomputed layout-aligned template cache to avoid repeated rendering. Empirically, across four public smart-home datasets under near-realistic mixed-activity protocols, LastAct achieves competitive or superior performance on pure windows and yields substantial Macro-F1 gains on cross/mixed windows, demonstrating improved robustness under near-realistic sliding-window regimes.

Multi-party dialogue discourse parsing aims to identify dependency structures and relation types between utterances in conversations. Previous studies are mostly limited to textual modality or two-party dialogue, failing to meet the multimodal and multi-party settings. In this paper, we construct the first publicly available English multimodal dataset DraDDP for multi-party dialogue discourse parsing, based on American TV dramas. DraDDP contains 495 dialogue segments with 6,374 utterances and 9.1 hours of parallel video content, covering rich multi-party interaction scenarios. Moreover, we establish comprehensive benchmarks by evaluating this task on DraDDP and conducting in-depth analysis on the impact of different modalities. Experimental results demonstrate the value of multimodal information in capturing dialogue structures and relation types. We will publicly release the dataset, annotation guidelines, and code to promote future research in multimodal dialogue understanding.

Unified multimodal models (UMMs) aim to handle perception and generation in a single model. Yet existing UMMs still rely on a frozen, separately pretrained VAE for image generation, imposing a structural bottleneck. Naively removing it introduces a quality gap, as the model must learn both high-level structure and low-level details from raw pixels. In this paper, we propose Representation Forcing (RF), a technique that closes this gap by making representation prediction a native capability of the model. Concretely, RF forces the decoder to autoregressively predict visual representations as intermediate tokens before pixels; these tokens then stay in context to guide pixel diffusion within the same backbone. By turning representations from perception outputs into generation targets, RF eliminates the need for any external generative latent space. We find that RF benefits both understanding and generation. On image generation, our pixel-space model with RF matches state-of-the-art VAE-based unified models. On image understanding, pixel-space RF generally outperforms its VAE-based variant. Together, these results offer an effective step toward end-to-end, bottleneck-free UMMs.

Despite Bengali being the sixth most spoken language in the world, no prior work has systematically evaluated hallucination in large language models (LLMs) for Bengali. We introduce BenHalluEval, a fine-grained hallucination evaluation framework for Bengali covering four tasks: Generative Question Answering (GQA), Bangla-English Code-Mixed QA, Summarization, and Reasoning. We construct 12,000 hallucinated candidates using GPT-5.4 across twelve task-specific hallucination types, drawn from three existing Bengali datasets, and evaluate seven LLMs spanning reasoning-oriented, multilingual, and Bengali-centric categories under a dual-track protocol that independently measures false-positive rate on ground-truth instances (Track A) and hallucination detection rate on hallucinated candidates (Track B). To jointly penalise both failure modes and prevent inflated scores from uniform response bias, we propose BenHalluScore, a dual-track calibration metric that ranges from 7.72% to 55.42% across models and tasks, revealing substantial variation in hallucination calibration. Chain-of-thought prompting, applied as a mitigation strategy, shifts response distributions without consistently improving hallucination discrimination. BenHalluEval establishes the first dedicated hallucination benchmark for Bengali and highlights the inadequacy of single-track and prompting-only evaluation approaches for low-resource language settings. The dataset and code are available at https://anonymous.4open.science/r/BanglaHalluEval-EB77.

Public consultations generate large volumes of data in the form of stakeholder submissions that are practically unfeasible to analyse manually. We present an end-to-end LLM-based pipeline and interactive dashboard for structured topic extraction from regulatory consultation submissions, demonstrated on the European Commission's Digital Fairness Act (DFA) public call for evidence as a case study. The system processes raw PDF attachments and web-form responses, extracts topic annotations, and grounds every extraction in a verbatim quote from the source text. Applied to 4,322 DFA submissions, the pipeline produced 15,368 topic annotations supported by 20,951 verbatim evidence quotes. Three principles govern the proposed design: verbatim grounding, full traceability, and transparency by design. The dashboard exposes the full extraction dataset through five analytical views, from dataset-level topic overviews to individual paragraph drill-downs, with every result traceable to its source. Beyond the predefined DFA topic categories, the pipeline generated certain stakeholder concerns, such as Age Verification, Payment Processor Censorship, and Digital Ownership, that a fixed-taxonomy approach would have missed. The pipeline is domain-generic; adapting it to a new consultation requires only a prompt update and a new dataset. A live demo is available at https://dfa-dashboard.thalesbertaglia.com/. The code and processed data are publicly available at https://github.com/thalesbertaglia/dfa-dashboard.

Regional accent classification in Brazilian Portuguese (pt-BR) suffers from the need for reliable labeling. While large self-supervised learning (SSL) speech models are powerful, their training pipelines dilute sociophonetic information, since accent labels are generally not reliable or are not used in training objectives. This work introduces a novel workflow for feature extraction using only acoustic labels. By isolating explicit regional accent landmarks and using a phoneme-based forced aligner (ZIPA), our targeted feature set captures dialectal variance more effectively than utterance embeddings, demonstrating that localized features can outperform general-purpose architectures on accent-related tasks using minimal and objective data labels.

Drawing on Ullmann-Margalit's concept of opting (transformative, irrevocable, and shadowed by foreclosed alternatives), we show that current AI systems raise a profound ethical problem that existing AI ethics has not fully captured: the illusion of opting, in which persons and groups encounter the deceptive appearance of meaningful consequential choice while the agency needed to become genuinely capable of choosing is weakened. Against approaches that treat AI primarily as an optimizer of already given ends, we argue that AI systems should be evaluated by whether they protect and cultivate meta-capacity against the illusion of opting: the socially and institutionally scaffolded agentive capacity through which means and ends can be formed, contested, revised, and owned. This reframing is especially urgent for disadvantaged populations, who are least able to absorb the costs of the illusion of opting when AI-mediated pathways misdirect behavior and action. We propose three normative imperatives for AI-mediated consequential decisions: existential honesty, which acknowledges the limits of prediction; ecological rationality, which situates guidance within heterogeneous lived ecologies; and counterfactual reparation, which acknowledges and repairs foreclosed alternatives when AI-mediated decision-making pathways fail.

Estimating individual treatment effect (ITE) from observational graph data is crucial for decision-making in the fields such as commerce and medicine. This task is challenging due to interference, where individual outcomes can be influenced by the treatments and covariates of their neighbors. Existing methods attempt to model such interference for accurate ITE estimation. However, a critical issue is often overlooked: differentiated networked effect (DNE), an effect caused by local networks consisting of neighbors with varying importance and scales. Capturing DNE is vital; otherwise, we will end up with imprecise ITE estimation due to an erroneous characterization of interference, which can result in misguided decisions. To address this challenge, we propose a novel interference modeling mechanism that incorporates two partial attention mechanisms and a message amplifier. The partial attention mechanisms automatically estimate the importance of different neighbors in contributing to interference, while the message amplifier adjusts the results of the interference modeling mechanism based on the scale of neighbors, all of which enables the model to capture DNE. Experiments on three real-world graphs demonstrate that our methods outperform existing approaches for ITE estimation from graph data, which corroborates the importance of explicitly capturing DNE.

Agentic systems increasingly run user-authored orchestration code that invokes tools, spawns subtasks, and delegates work across machines and clouds. Although this high agency is productive, it creates a security problem: identity, authorization, provenance, and delegation are often pushed into application code, where they become difficult to enforce consistently and difficult to audit. We present Grimlock, an Agent Guard that restores separation of concerns by moving trust enforcement into the sandbox substrate while leaving agent code unchanged. Grimlock uses eBPF-enforced traffic interception to ensure that sandbox communication passes through a guard, and combines it with post-handshake attestation bound to standard TLS~1.3 channel bindings. After a channel is established, the guard authorizes communication and mints short-lived, channel-bound scope tokens that capture least-privilege delegation. At the receiving side, the destination guard re-validates identity, scope, and channel binding, terminates TLS, and releases plaintext to the destination sandbox only after policy checks succeed. kTLS provides an efficient dataplane for protected communication. As a result, Grimlock offers a path toward transparent, auditable, and scope-bound agent-to-agent communication across heterogeneous multi-cloud environments, using commodity Linux primitives and without requiring changes to user-layer orchestration code.

We train a pair of autoregressive models to construct zero-mean control variates to mitigate the sign problem in quantum Monte Carlo simulations. The two autoregressive networks are confined to the positive- and negative-sign sectors with strictly disjoint support, and each is exactly normalized over its sector. Their difference is therefore structurally zero-mean, providing an unbiased auxiliary observable whose correlation with the sign estimator controls the variance reduction. We implement the method within the stochastic series expansion framework, which we extend to frustrated lattices by developing an incremental loop-topology update. Sign-ergodic sampling is achieved through a twist channel, which is the unique sign-changing mechanism on non-bipartite lattices. We implement the control variates as autoregressive transformers with an end-of-sequence parity mask that enforces exact sign-sector resolution, while the incremental loop-count change and cumulative frustration parity are incorporated as topological features. On the triangular-lattice Heisenberg antiferromagnet, we benchmark the method in the small-$N$ limit. The control variate reduces the standard error of the average sign by up to an order of magnitude and that of the energy estimator by a factor of three to five, remaining effective even when the average sign drops below $10^{-3}$. This work lays out the framework and provides a proof-of-principle demonstration that autoregressive control variates can effectively mitigate the sign problem. Scaling to larger systems with physics-informed architectures is the subject of future work.

Multi-vector retrieval (MVR) models, exemplified by ColBERT, have established new benchmarks in retrieval accuracy by preserving fine-grained token-level interactions. However, this granularity imposes prohibitive storage and retrieval efficiency bottlenecks: to manage the immense memory footprint and computational overhead of billion-scale token vectors, state-of-the-art systems are forced to rely on aggressive dimension reduction and complex clustering (e.g., K-means). This compromise introduces two critical limitations: excessive indexing latency of clustering large-scale corpora and semantic information loss inherent to compression. In this paper, we propose Single-stage Sparse Retrieval (SSR}, a paradigm shift that replaces expensive clustering with efficient sparse coding. Instead of compressing features into low-dimensional dense vectors, we utilize Sparse Autoencoder (SAE) to project token embeddings into a high-dimensional but highly sparse representation. This transformation enables us to bypass vector clustering entirely and leverage inverted indexing for precise, high-throughput retrieval. Extensive experiments on the BEIR benchmark demonstrate that SSR achieves a "trifecta" of improvements: it reduces indexing time by 15x compared to ColBERTv2, halves retrieval latency, and simultaneously improves retrieval performance over leading baselines.

Many open-ended instructions have multiple valid answers that users can benefit from seeing, but post-training often narrows an LLM's output space toward a small set of canonical responses. We introduce REDIPO, an offline DPO data-construction pipeline for recovering distinct valid answer modes while preserving the alignment benefits of the instruct model. For each prompt, REDIPO samples responses from both base and instruct models, rewrites base-model responses with the instruct model, filters candidates for safety and instruction-following quality, and builds preference pairs that favor marginally diverse responses among candidates with similar instruction-following reward. Across Qwen3-4B, OLMo-3-7B, and LLaMA-3.1-8B, REDIPO improves NoveltyBench distinct_k by 134%, 33%, and 44% relative to the instruct checkpoints, while DivPO changes diversity by 0%, -6%, and -4% on the same models. These gains largely maintain MTBench, IFEval, and Arena-Hard performance, and reduce direct-category HarmBench attack success rate. Ablations show that marginal-diversity pair selection and base-response rewriting drive the diversity gains, while filtering and quality-bounded pairing help maintain alignment. Overall, our results show that diverse valid answers from base-model generations can be reintroduced through carefully constructed preference data while retaining the alignment benefits of post-training. We release our code and data at https://github.com/vsamuel2003/ReDiPO.

As AI-generated and AI-assisted content floods online spaces, source labels attached to such content can distort human reasoning judgments, with downstream consequences for moderation, evaluation, and decision-making. Whether LLMs share this vulnerability, or offer more source-agnostic evaluation, remains an open question with direct implications for human-AI collaboration. We examine this issue using logical fallacies as a controlled setting to isolate source-label effects on reasoning quality, independent of domain knowledge. We conduct an online study (N=505) where participants are assigned to a source condition (human, AI, human with AI assistance, AI with human assistance, or no disclosure) and evaluate comments containing logical fallacies, comparing their judgments with those of LLMs (GPT-5.2, Gemini 2.5 Flash, Claude Sonnet 4.5), who were evaluated across the same source conditions. Human evaluators were significantly more susceptible to fallacies labeled as written by human or human with AI assistance and assigned higher trust and evaluation ratings in these conditions. LLM evaluations remained comparatively stable across source labels, though performance varied across models. Confidence levels were similarly high across conditions for both humans and LLMs, regardless of fallacy presence. Our findings indicate that source-label bias in reasoning evaluation is primarily a human vulnerability and highlight the potential of human-LLM collaboration in increasingly AI-mediated environments.

Large Language Models (LLMs) have advanced autonomous agents from deep search, which retrieves concise factual answers, to deep research, which synthesizes scattered evidence into long-form reports. However, verifiable multimodal deep research remains challenging due to open-ended synthesis without deterministic ground truth and the need to interleave textual arguments with visual evidence. We propose Ptah, a multi-agent harness for interleaved report generation. Ptah orchestrates the lifecycle from user query to rendered web report through planning, research, and writing stages, where specialized agents construct visual-aware plans, collect claim-grounded evidence, maintain source-aligned images in a Visual Working Memory, and compose reports through declarative multimodal tool use. A verifier agent serves as the harness's acceptance function, enforcing factual grounding, citation fidelity, and cross-modal consistency throughout the workflow. We further introduce PtahEval, an evaluation protocol that augments existing benchmarks with image-level and presentation-level assessments. Experiments on deep research benchmarks show that Ptah produces more reliable, visually informative, and usable human-facing multimodal reports than strong baselines. Our code is released at https://github.com/SnowNation101/Ptah

Reinforcement learning (RL) is a natural fit for agentic knowledge base question answering (KBQA), where a model must issue executable actions, observe knowledge-base feedback, and eventually return an answer. However, current RL-based KBQA systems mainly optimize sparse rewards from the final answer, leaving intermediate action errors weakly supervised. This is especially limiting for logical-form annotated KBQA benchmarks: gold logical forms can be converted into executable action sequences, but existing pipelines use them mainly for warm-start data construction rather than for on-policy RL updates. We propose GAPD, a training-time Gold-Action Policy Distillation framework that adds dense token-level guidance to outcome-based RL. To align gold actions with on-policy student rollouts, GAPD uses MID-ANCHOR MATCHING: it treats the intermediate entities reached during student exploration and gold execution as state anchors, and matches student states to gold states through these explored entity sets. The current policy conditioned on this aligned gold action serves as a stop-gradient teacher, whose token distribution is distilled back to the ordinary student policy over generated action-token spans. GAPD consistently surpasses the current state of the art on WebQSP, GrailQA, and GraphQ.

Estimating probability density and its score from samples remains a core problem in generative modeling, Bayesian inference, and kinetic theory. Existing methods are bifurcated: classical kernel density estimators (KDE) generalize across distributions but suffer from the curse of dimensionality, while modern neural score models achieve high precision but require retraining for every target distribution. We introduce DiScoFormer (Density and Score Transformer), a ``train-once, infer-anywhere" equivariant Transformer that maps i.i.d. samples to both density values and score vectors, generalizing across distributions and sample sizes. Analytically, we prove that self-attention can recover normalized KDE, establishing it as a functional generalization of kernel methods; empirically, individual attention heads learn multi-scale, kernel-like behaviors. The model converges faster and achieves higher precision than KDE for density estimation, and provides a high-fidelity plug-in score oracle for score-debiased KDE, Fisher information computation, and Fokker-Planck-type PDEs.

Search agents connect LLMs to the Internet, enabling them to access broader and more up-to-date information. However, this also introduces a new threat surface: unreliable search results can mislead agents into producing unsafe outputs. Real-world incidents and our two in-the-wild observations show that such failures can occur in practice. To study this threat systematically, we propose SafeSearch, an automated red-teaming framework that is scalable, cost-efficient, and lightweight, enabling sandboxed safety evaluation of search agents. Using this, we generate 300 test cases spanning five risk categories (e.g., misinformation and prompt injection) and evaluate three search agent scaffolds across 17 representative LLMs. Our results reveal substantial vulnerabilities in LLM-based search agents, with the highest ASR reaching 90.5% for GPT-4.1-mini in a search-workflow setting. Moreover, we find that common defenses, such as reminder prompting, offer limited protection. Overall, SafeSearch provides a practical way to measure and improve the safety of LLM-based search agents.

Existing Learning-to-Defer (L2D) frameworks are limited to single-expert deferral, forcing each query to rely on only one expert and preventing the use of collective expertise. We introduce the first framework for Top-$k$ Learning-to-Defer, which allocates queries to the $k$ most cost-effective entities. Our formulation unifies and strictly generalizes prior approaches, including the one-stage and two-stage regimes, selective prediction, and classical cascades. In particular, it recovers the usual Top-1 deferral rule as a special case while enabling principled collaboration with multiple experts when $k>1$. We further propose Top-$k(x)$ Learning-to-Defer, an adaptive variant that learns the optimal number of experts per query based on input difficulty, expert quality, and consultation cost. To enable practical learning, we develop a novel surrogate loss that is Bayes-consistent, $\mathcal{H}_h$-consistent in the one-stage setting, and $(\mathcal{H}_r,\mathcal{H}_g)$-consistent in the two-stage setting. Crucially, this surrogate is independent of $k$, allowing a single policy to be learned once and deployed flexibly across $k$. Experiments across both regimes show that Top-$k$ and Top-$k(x)$ deliver superior accuracy-cost trade-offs, opening a new direction for multi-expert deferral in L2D.

Large Language Models excel in generative tasks but exhibit inefficiencies in structured text selection, particularly in extractive question answering. This challenge is magnified in resource-constrained environments, where deploying multiple specialized models for different tasks is impractical. We propose a Learning-to-Defer framework that allocates queries to specialized experts, ensuring high-confidence predictions while optimizing computational efficiency. Our approach integrates a principled allocation strategy with theoretical guarantees on optimal deferral that balances performance and cost. Empirical evaluations on SQuADv1, SQuADv2, and TriviaQA demonstrate that our method enhances answer reliability while significantly reducing computational overhead, making it well-suited for scalable and efficient EQA deployment.

Knowledge distillation (KD) transfers a single scalar prediction from a large foundation model (FM) to compact vertical models (VMs), suffering from diminishing transfer ratio -- the fraction of FM improvement captured by the VM -- as a single scalar cannot convey the rich intermediate knowledge that larger FMs learn. To address this bottleneck, we propose LoopFM (Learning frOm HistOrical RePresentations of FM), a framework that opens a high-bandwidth transfer channel by structuring FM intermediate embeddings as input features (e.g., user history sequence) for downstream VMs, without requiring real-time FM inference at serving and architectural coupling between FM and VM. We provide a theoretical framework for LoopFM with a gain decomposition and transfer-ratio analysis. On three public benchmarks, LoopFM demonstrates strong AUC improvements (e.g., 6%+ on TaobaoAd) and complementary knowledge transfer capability with KD. On industrial-scale systems (billions of examples, trillion-parameter FMs), LoopFM approximately doubles the knowledge transfer ratio on top of KD, delivering a +0.5% conversion improvement in the first half after its initial launch, and +1.03% and +1.22% conversion improvement from two individual launches in the subsequent half.

LLMs have advanced text classification, yet existing paradigms face a trade-off: supervised (label only) fine-tuning is scalable but offers limited reasoning on complex text and lacks broader model transparency, while discrete prompt optimization offers human-readable instructions but struggles with performance and scalability. We introduce eXTC (eXplainable Text Classifier) with three progressive stages: (1) learning a Standard Operating Procedure (SOP, or rulebook) in natural language via a new Structured Prompt Optimization algorithm; (2) SOP-grounded reasoning distillation from a large teacher LLM into a compact LM; and (3) expanding reasoning capabilities beyond the initial SOP via reinforcement learning. This design enables eXTC to provide (i) fast inference via a compact LM, with (ii) inference-time local reasoning traces, alongside a global, modular explanation of its learned domain rules, while (iii) significantly outperforming existing paradigms across diverse benchmarks in both classification performance and explanation quality, with stage-by-stage gains.

Competitive STEM examinations such as JEE and NEET require multi-step symbolic reasoning, precise numerical computation, and deep conceptual understanding across physics, chemistry, and mathematics. Recent large language models perform strongly on common reasoning benchmarks, yet they remain difficult to deploy at scale, where millions of student doubts demand domain-specific, consistently structured problem solving. We introduce Aryabhata 2, a reasoning-focused language model for competitive STEM examinations, trained via reinforcement-learning post-training. Using PhysicsWallah's internal question banks, we construct a high-quality training curriculum and post-train GPT-OSS-20B through reinforcement learning with verifiable rewards. Training combines prolonged reinforcement learning with broadened exploration via progressively larger rollout group sizes. We evaluate Aryabhata 2 on competitive examination benchmarks, including JEE Main, JEE Advanced, and NEET, as well as out-of-distribution reasoning datasets such as AIME, HMMT, MMLU-Pro, MMLU-Redux 2.0, and GPQA. Results show that Aryabhata 2 outperforms its base model GPT-OSS-20B on competitive STEM reasoning while requiring substantially fewer output tokens (up to 64\% fewer).

Self-supervised pre-training paradigm has gained increasing prominence for learning transferable representations in medical imaging, yet existing methods for ultrasound (US) images operate at the image or frame level, overlooking the anatomical context for clinical-aligned representation learning. In this work, we propose an anatomy-anchored ultrasound self-supervision framework ANAUS that shifts representation learning from generic visual regions to clinically meaningful anatomical structures. Utilizing a learnable latent prompt engine alongside a one-time domain adaptation on existing public image-mask pairs, we empower the LP-SAM module to achieve annotation-free anatomy delineation at scale. Building upon this anatomical grounding, we propose a dual-policy self-supervised learning paradigm consisting of inter-view semantics-aware anatomy-separating alignment and contextual core-region prediction to enhance representation learning. Specifically, the former enforces feature invariance within identical anatomical regions while promoting discriminability across distinct structures; the latter compels the model to reconstruct corrupted regions, thereby capturing fine-grained structural details. Extensive evaluations on six public datasets demonstrate that ANAUS consistently outstrips current state-of-the-art methods while maintaining the computational efficiency essential for clinical deployment. Code is available at https://github.com/zhcz328/ANAUS.