论文库

We present Meituan-Robotics-0 (MR0), a vision-language-action foundation model and embodied-agent system for stationary-base desktop manipulation. MR0 combines 147 million vision-language samples with 15,060 hours of curated robot trajectories spanning 12 embodiments. A three-stage training recipe aligns multimodal representations with robot behavior through discrete action-token pre-training, transfers these priors to continuous cross-embodiment control through flow-matching mid-training, and adapts the policy through real-world post-training. We introduce Rollback, which pairs replaced policy segments with human corrections to provide positive and negative supervision for CFG-DAgger without value learning or advantage labeling. A unified end-effector interface supports heterogeneous embodiments, while asynchronous action-chunk merging maintains smooth execution under variable inference latency. An event-driven agent maintains persistent task events and reassesses plans when progress or feasibility changes, enabling cross-task retrieval and conditional task resumption. MR0 achieves 93.6% success on RoboTwin 2.0 and ranks fifth in the RoboDojo VLA track. Under matched HG-DAgger post-training, MR0 achieves higher average success rate and progress score than $π_{0.5}$ across three real-robot tasks. On Ethernet insertion, CFG-DAgger trained with Rollback data reaches 70% success and a progress score of 0.90. Qualitative agent deployments demonstrate cross-task event retrieval and recovery from temporarily unsatisfied task conditions.

The Witness is a game with combinatorial puzzles that are challenging for both human players and AI solvers. Search-based AI solvers can require large amounts of time to solve such puzzles. We accelerate search by automatically learning a predicate that predicts whether a partial path to a Witness puzzle is completable to a solution path. We demonstrate that our learned predicate is human-explainable, which makes it amenable to proofs, and we prove a key property that allows us to use it for pruning successor states in search. Doing so accelerates search by about six times while maintaining search completeness. Conversely, given a fixed time budget per puzzle, our predicate-accelerated search can solve more puzzles of larger sizes than a baseline search using a hand-coded predicate. Finally, we empirically demonstrate how accelerating search using an automatically learned predicate can outperform two competitors, weighted A* and Levin tree search with neural networks, in terms of time and number of nodes expanded.

In many strategic settings, agents can influence opponents by selectively disclosing truthful information about their own future actions. We introduce a game-theoretic framework to model this phenomenon. As a case study, we consider a two-player, alternating-move epistemic game in which Player A evaluates actions under a conservative maximin rule, while Player B-the manipulator-leverages this behavior by strategically revealing commitments. We prove that the manipulator can implement every inducible outcome using at most n disclosures in a binary game of depth 2n with n decisions per player. This establishes a sharp contrast between the game tree's exponential size and the linear amount of information required to control the outcome, highlighting a general mechanism by which limited information can exert disproportionate strategic influence.

We study list update with time windows and delays when every request requires a separate access. Unlike batch-serving models, serving several requests to one element still incurs one access per request, and each access may change the list for subsequent requests. For time windows of overlap at most $k\ge2$, our deterministic First-Element Removal algorithm is $(3-2/k)$-competitive, hence $3$-competitive in general. Against an oblivious adversary, a randomized adaptation of BIT is $7/4$-competitive on arbitrary $2$-overlap inputs. More generally, on inputs without self-overlap, it is $\left(13/4-9/[2(k+1)]\right)$-competitive for every $k\ge2$. For delays, we give a deterministic $5$-competitive algorithm and a lower bound of $2+\sqrt3$ for deterministic delay-unaware algorithms. Our analyses extend inversion potentials with per-request credits that reconcile different service times and orders.

Recent work in the privacy literature shows that sample-targeted membership inference attacks (MIAs) significantly outperform untargeted approaches by a wide margin. Motivated by this observation, we address the following question: can the privacy vulnerability of individual training points be assessed without training shadow models? We show that per-sample exposure to MIA is governed not only by a point's loss, but also by a data-dependent geometric measure. In the linear setting, we derive a closed-form decomposition of individual black-box MIA vulnerability into a population leverage score and a residual loss term, making explicit how sample-dependent geometry translates into privacy exposure. Since the final layer of most modern architectures is linear, we extend this framework to deep networks and propose a surrogate score operating on last-layer representations that requires only a single trained model and no shadow models. Empirical evaluations across diverse datasets and architectures show that our score outperforms loss and gradient-norm baselines at identifying the highest-risk points under state-of-the-art attacks, providing a computationally efficient and theoretically grounded tool for per-sample privacy risk assessment.

Change detection in remote sensing conventionally relies on spatially aligned bi-temporal imagery, which may be unavailable or unreliable in practical scenarios. Recent single-temporal approaches address this limitation by learning structural evidence, such as building regions, from a labeled source dataset and transferring it to unseen target imagery using only the current image. However, adaptation to the source dataset can over-specialize pretrained features and reduce their cross-dataset generalization, a risk that this paper argues is higher when the feature adaptation is parameterized independently of the states used for prediction. To address this issue, the current study introduces SALT, a State-Aligned Low-Rank Transfer framework. SALT keeps a DINOv3 backbone frozen and learns negative and positive structural states that both score each patch and, through State-Aligned Low-Rank Adaptation (SALA), span the rank-2 subspace in which patch features are corrected. Directional State Alignment supervises these corrections to raise the evidence of positive patches and lower that of negative ones, while Geometry Preservation limits changes to the pairwise relations among pretrained patch features. Experiments on LEVIR-CD and the WHU aerial building dataset show that, at the pixel level and in both transfer directions, SALT outperforms four baselines reproduced under a common source-only protocol, improving over the strongest reproduced baselines by 9.45 and 5.79 points in pixel-level AP and by 27.69 and 14.25 points in IoU (LEVIR-CD to WHU and WHU to LEVIR-CD, respectively), with only 0.068M trainable parameters and 40.2 ms per image. In order to facilitate reproducibility and further research, the source code of SALT is available at https://github.com/gthpapadopoulos/SALT

We investigate the minimal width of deep neural networks that can approximate all continuous functions from $\mathbb{R}^n$ to $\mathbb{R}^m$ in the topology of uniform convergence on compact subsets. We show that whenever the real projective space $\mathbb{RP}^{n-1}$ admits a smooth immersion into $\mathbb{R}^n$, the minimal width is at least $n+\min\{n,m\}$, assuming that the activation function is continuous and monotone. As a consequence, the minimal width of ReLU networks is $n+m$ whenever $n$ is equal to 2, 3, 4, 7 or 8 and $1\le m \le n$, extending results of Hanin-Sellke and Hwang.

Ontology matching (OM) enables semantic interoperability between different ontologies and resolves their conceptual heterogeneity by aligning related entities. OM systems currently have two prevailing design paradigms: conventional knowledge-based expert systems and newer machine learning-based predictive systems. While large language models (LLMs) and LLM agents have revolutionised data engineering and have been applied creatively in many domains, their potential for OM remains underexplored. This study introduces a novel agent-powered LLM-based design paradigm for OM systems. With consideration of several specific challenges in leveraging LLM agents for OM, we propose a generic framework, namely Agent-OM (Agent for Ontology Matching), consisting of two Siamese agents for retrieval and matching, with a set of OM tools. Our framework is implemented in a proof-of-concept system. Evaluations of three Ontology Alignment Evaluation Initiative (OAEI) tracks over state-of-the-art OM systems show that our system can achieve results very close to the long-standing best performance on simple OM tasks and can significantly improve the performance on complex and few-shot OM tasks.

We study community search when each member has its own tolerance for differences in position, as when members of a group interact repeatedly and influence one another. A latitude-consistent community (LCC) contains the query vertex, is connected, has minimum degree at least k, and bounds its diameter by the smallest member tolerance. The bound is exactly the condition under which every interaction in a connected closed group stays accepted under bounded-confidence dynamics for every choice of positive weights. It also prevents rejection under social-judgment dynamics, whereas the pairwise condition and other natural relaxations preserve neither guarantee. Finding a maximum LCC is NP-hard even in the plane, and with unbounded dimension, deciding whether an LCC exists is NP-complete for every fixed k>=2. LatSweep fixes the answer's smallest tolerance and turns the coupled bound into a best-first sweep over candidate thresholds, yielding an exact branch-and-bound algorithm, a (1,2)-bicriteria approximation and two feasible heuristics. Across eight graphs with up to 1.8 billion edges, six with synthetic positions and tolerances, the exact algorithm proves optimality on all but 9 of 3,700 efficiency queries; in a shared implementation on two large graphs, its branch-and-bound solver proves more queries than a reimplemented maximum-(k,r)-core algorithm, a mixed-integer program and CP-SAT. In simulations of closed groups, LCC groups keep every accepted interaction under uneven influence, whereas 3-40% of pairwise-compatible groups lose some. On three location-based networks, tolerance-agnostic groups often place another member's home beyond a member's estimated travel radius. Enforcing every member's tolerance causes no measurable loss of co-visits relative to spatial communities of the same size. Relative to the larger pairwise-compatible groups, it can cause a size-adjusted loss.

Fine-tuning foundation models on financial data has a wide range of applications, including identifying event types from news, monitoring risk and compliance language, and answering questions over filings and earnings calls. In this setting, the main challenge is to improve task performance without degrading the generalization capabilities of the model. In this paper, we present an extensive set of ablations for fine-tuning Qwen3-8B with low-rank adaptation (LoRA). We compare three training regimes: (1) response-only LoRA, where the supervised response is the label alone; (2) single-adapter reasoning, where reasoning traces from a teacher model are included alongside the labels in the training loss; and (3) dual-adapter composite training, where separate LoRA adapters are trained on traces and labels. Response-only LoRA improves in-domain performance and leaves general capabilities near those of the frozen model. The in-domain change is largest for calibration and argmax accuracy. Under single-adapter reasoning, the in-domain metrics remain behind those of the response-only regime, while MMLU-Pro can rise slightly and the other retention benchmarks stay within ordinary variation. Dual-adapter composite training remains on the same side of this comparison and yields no further gain. Finally, we compress the teacher traces by summarizing them before training, and show that shorter summaries move calibration toward the response-only regime without reaching it.

Automatic setlist identification aims to identify the songs performed in live concert recordings by matching noisy and variable-quality audio against a database of reference tracks. This task presents challenges beyond conventional music version identification, including substantial arrangement differences, background noise, and the need to process continuous recordings efficiently at large scale. In this work, we evaluate modern music version retrieval methods for setlist identification using the ASID benchmark, comparing the established Re-MOVE system with CLEWS and NAVE, an internal neural audio version embedding model. Our results show that NAVE and CLEWS substantially outperform Re-MOVE. Notably, NAVE uses 256-dimensional embeddings, compared with 1024 dimensions for CLEWS, providing a four-fold reduction in embedding storage while achieving better retrieval performance. We further evaluate music/non-music gating and find modest improvements in detection and temporal coverage, while qualitative analysis suggests that such preprocessing can improve recognition boundaries beyond what the standard ASID metrics capture. Overall, the results demonstrate that segment-level audio embeddings provide a strong basis for scalable setlist identification and that compact representations can achieve strong accuracy under realistic live-concert conditions at much lower storage and serving cost.

Program evolution provides a promising approach to discover agent memory designs and study the choices underlying their effectiveness. We introduce MAPE, a framework that evolves whole memory programs from minimal seeds through a standardized lifecycle interface and an iterative diagnosis-proposal loop. Experiments on LoCoMo and LongMemEval yield an absolute accuracy gain of up to 12% over manually designed baselines at 35% of the serving cost, with further evaluations supporting transfer across serving backbones and benchmarks. Empirical analyses and ablations identify and validate two recurring design patterns: retaining raw conversations as retrievable evidence preserves details omitted from extracted memories, while composing retrieved evidence according to query intent helps the answering model use it effectively. Transplanting the composition module into five existing memory systems improves accuracy across all five. Beyond these patterns, the search objective shapes where computation goes: optimizing serving cost alongside accuracy shifts computation from LLM-based memory construction to selective processing at query time while retaining competitive accuracy. Together, these findings offer empirical lessons for agent memory design. Code is available at https://github.com/ash-ding/MAPE.

System thermal-hydraulics codes such as the System Analysis Module (SAM) close turbulence on coarse meshes with algebraic zero-equation mixing-length models. These models cannot represent the non-equilibrium turbulence, flow separation, and large-scale recirculation of open-cavity advanced reactors. We develop an invariant neural-network eddy-viscosity closure for SAM and apply it to the Molten Salt Fast Reactor (MSFR). The closure predicts the turbulent kinetic energy and timescale of each element from the flow and geometry of its neighbors, without solving turbulence transport equations. The resulting non-negative eddy viscosity enters the implicit diffusive operator, which keeps the momentum equations well conditioned. We prove that the closure is exactly invariant to translation, Galilean boost, rotation, reflection, and neighbor ordering. A network trained on high-order Nek5000 fields sees a different feature distribution on the coarse SAM mesh. We remove this covariate shift by iteratively retraining the network on SAM flow states against fixed Nek5000 targets. For a held-out steady state, the coupled closure reproduces the single recirculation cell and eddy-viscosity plume of the Nek5000 reference. In a 120 s pump-coastdown transient, it halves the velocity error of the mixing-length model, from 37.6% to 17.7%, and matches the Nek5000 eddy viscosity with a time-averaged relative error of 12%. The remaining velocity error comes from the coarse mesh. The coupled transient takes about 0.2 core-hours, compared with about 330 core-hours for Nek5000. The results show that the closure can recover two-equation RANS turbulence behavior in a coarse-mesh system code at a small fraction of the CFD cost.

FPGAs are widely used in mission-critical systems, where a successful physical attack can have serious consequences. This makes resilience to such attacks a critical security requirement. To defend against active physical attacks, such as glitching attacks, modern FPGAs rely on on-chip sensors that detect these disturbances and trigger responses, such as key zeroization, before a device is compromised. These sensors are deployed as both hard-IP blocks and soft-IP cores, and their trustworthiness is often taken for granted. In this work, we challenge that assumption and argue that the sensors themselves are a fragile and underexplored attack surface. Drawing on recent results, we show that both the detection and response mechanisms of state-of-the-art sensors can be bypassed without invasive tampering. We further argue that sensor reliability is reduced by factors rarely captured in threat models, including the spatial and frequency-dependent behavior of the system's power delivery network, which can steer a fault to a target while sparing nearby sensors; environmental conditions such as temperature, which shift detection thresholds; and long-term aging, which drifts sensor calibration and may open silent detection gaps. Finally, we propose a set of countermeasures that harden detection and response mechanisms against these effects.

Large foundation models such as GPT-6 Astra and Claude Fable 5.1 have recently been shown to control robots directly from visual observations. They nevertheless remain brittle on long-horizon, contact-rich tasks in real-world environments. Further progress depends on two complementary resources: demonstrations of human behavior that supervise how tasks are performed, and interactive environments in which policies can be rolled out, evaluated, and improved in closed loop. Language models have advanced along well-characterized scaling laws; the corresponding question for physical AI is how capability scales with environments and behaviors, which first requires identifying which data to scale and how to produce and reuse such data continuously. To address this question, we introduce Ropedia Xperience 2, comprising 1,589 released metric-scale 3D scans and a collection of 4,230 egocentric recordings of 74 tasks (131 hours). We further show that each capture can be repurposed for robot learning: scenes are converted into physically interactive simulation environments, human motion is retargeted to whole-body robot trajectories, and additional robot behaviors are obtained through virtual teleoperation and automated long-horizon trajectory generation. We propose this joint capture of environments and behaviors as a gold standard for physical AI data collection, laying a data foundation for establishing scaling laws over environments and behaviors.

Online conformal risk control (OCRC) enables reliable uncertainty quantification in machine learning by converting model outputs into prediction sets with long-term risk guarantees. These guarantees, however, hinge on the assumption that the loss of each round is revealed immediately and losslessly. In practice, feedback may be delayed, and limited resources may prevent tracking all pending predictions, causing some feedback to be lost. To address this challenge, we develop capacity-aware OCRC (CA-OCRC), a framework that jointly determines how to allocate limited tracking feedback capacity and how to construct prediction sets. CA-OCRC is based on a randomized tracking policy that samples a proxy delay to decide whether, and for how long, feedback for the current round is tracked. CA-OCRC updates the prediction sets via inverse probability weighting, compensating for the possibility of feedback lost due to delays and limited tracking capacity. We analyze risk bounds for both arbitrary and stochastic delays, revealing a trade-off between receiving more feedback and avoiding capacity saturation. Building on these bounds, we optimize the proxy delay distribution, showing scaling laws for the tracking capacity that ensure the same long-term risk as OCRC with unlimited feedback. Experiments on medical imaging, natural language processing, computer vision, and wireless communications show the effectiveness of CA-OCRC as compared to existing OCRC mechanisms.

Human-AI interaction today runs almost entirely through prompts and outputs, keeping AI systems opaque to general users. We introduce Neural Representation Interaction (NRI), a paradigm that treats the internal representations of neural networks as an interaction material people can inspect, steer, and edit. Situating NRI alongside graphical, tangible, and bio-computer interaction, we argue that the concepts and algorithms models encode internally offer representational affordances for richer human-AI interaction. We synthesize actionable mechanistic interpretability methods and, drawing on General Morphological Analysis, construct a design space spanning Interaction, Target, Referent, and Purpose. Mapping existing interfaces onto this space reveals that most serve expert auditors, while few support lay users. We then present speculative applications in education, personal user modeling, and hiring audits, and discuss design principles, limitations, and sociotechnical risks. NRI offers HCI researchers abstractions for designing interactions that work with AI computation rather than only through its inputs and outputs.

Large language models (LLMs) have been widely adopted for clinical question answering (QA). Current systems can attach citations to their answers, but these often point to broad texts, leaving time-pressed clinicians unable to verify them efficiently. An alternative is to ensure that responses are verifiable by construction: providing fine-grained verbatim quotes from reference material that substantiate claims, so users can verify an answer without opening other documents. In this paper, we evaluate the ability of current models to perform this task end-to-end: from providing citations for every factual claim, to producing verbatim quotes, to ensuring that those quotes fully substantiate the claims. To do so, we build a standardized harness over four clinical practice guidelines and evaluate fourteen LLMs on 222 synthetic clinical questions, measuring each of these stages separately. We find that most models can attach verbatim quotes to over 90% of their claims from prompting alone, apart from some lightweight models such as claude-haiku-4.5. Yet these quotes often fail to substantiate every detail of the claims they accompany. For instance, claude-opus-5 produces verbatim quotes for 98.0% of its claims, but fully substantiates only 37.1%. Our work provides insights into the current capability gap of LLMs in building verifiable clinical QA systems, along with artifacts for future research.

Structured Electronic Health Records (EHR) models typically treat each clinical concept as a dataset-specific token, discarding the natural-language semantics associated with the event. As a result, relationships between concepts must largely be learned from EHR co-occurrence, without directly exploiting any prior semantic knowledge. We investigate how pretrained causal language models can be adapted for longitudinal EHR prediction by developing tEHR. Here, we represent events with their natural language description and use this text stream to apply EHR-adaptive continued pretraining, followed by task-specific discriminative fine-tuning using LoRA, utilising the LLM as a longitudinal patient-encoder. We evaluate this approach across four prediction tasks on the MIMIC-IV dataset and pancreatic cancer prediction on the CPRD dataset at 1 to 12 month horizons, with tEHR using the Mistral-7B-v0.3 backbone achieving the strongest overall performance. Ablations show that performance declines when clinical descriptions are removed or misaligned, when numerical values are omitted, and when either stage of adaptation is removed, while explicit elapsed-time descriptions provide little additional benefit. Analysis of the learned patient representations further reveals clinically coherent structure associated with distinct patterns of disease risk and clinical history.

LLM-based peer review is now used at production scale, but its evaluation remains limited. We introduce a benchmark that injects objective, localisable errors of four types (general issues, mathematics, numerics, citations) at three severity levels into real papers and measures each reviewer's detection rate by error type and severity, together with its false-positive rate on clean copies. We also present Skepthical, a multi-agent reviewer that verifies citation-backed claims by retrieving the cited papers, mathematics with computer algebra, and numerical claims with executed code. We apply the benchmark to two corpora of 15 recent arXiv preprints each, one in astrophysics and one in machine learning (ML), injecting a total of 357 errors, and evaluate twelve reviewer systems: six single-LLM reviewers, two Claude Code configurations, three publicly available reviewers, and Skepthical. The choice of model has the largest impact on detection rate: given the same prompt, three frontier LLMs differ by about 50 percentage points in total detection on both corpora. Tools and scaffolding help when they contribute what the model cannot supply on its own. In particular, on astrophysics, Claude Code improves on a single Claude call in mathematics and numerics detection, and enforced citation retrieval gives Skepthical the highest citation error detection rate of all systems by a large margin. This citation lead, however, is not observed on ML, where the models appear to know the literature better and Skepthical's citation detection rate is statistically indistinguishable from that of the best single LLM. False-positive rates are below 10% for most systems, but vary with prompt, model, and scaffolding, with mathematics flags the main source of false positives. The evaluation pipeline, the injected errors, and the adjudication logs are released as a benchmark.

We study interactive proofs of proximity (IPPs) for model evaluation, where a resource-limited verifier interacts with an untrusted prover, typically the model owner, to certify statistical properties of a model under an unknown input distribution. Our formulation separates sampling the input distribution from querying the model and evaluating its output; distinguishes real audit data (black-box sampling) from generated data (chosen-randomness, or gray-box, access to the sampler); and allows the prover and verifier to use different evaluators. We focus on doubly-sublinear IPPs, where both the verifier and honest prover use sublinear resources, and on (weighted) Hamming weight properties. For ordinary Hamming weight, we give a tolerant doubly-sublinear IPP. For completeness and soundness radii $\varepsilon_c<\varepsilon_f$ and gap $g=\varepsilon_f-\varepsilon_c$, a logarithmic-round instantiation uses $\widetilde{O}(1/g)$ verifier queries and $O(1/g^2)$ honest-prover queries, improving the cubic dependence of Amir, Goldreich, and Rothblum (ITCS 2025). We prove matching query lower bounds up to polylogarithmic factors. For distribution-weighted Hamming weight, black-box sampling requires $Θ(1/g^2)$ verifier samples but only $\widetilde{O}(1/g)$ evaluations; the quadratic sample complexity is necessary in the interior regime. With chosen-randomness access, the problem reduces to ordinary Hamming weight, yielding $\widetilde{O}(1/g)$ calls and evaluations. If the parties' evaluators disagree arbitrarily on a $ρ$-fraction of the distribution and by at most $γ$ elsewhere, our protocols remain doubly sublinear whenever $g>2κ$, where $κ=ρ+(1-ρ)γ$. Applications include auditing accuracy, group fairness, calibration, harmlessness, usefulness, and average-case robustness.

We give faster algorithms for diameter and farthest-neighbor search in high-dimensional Euclidean space. For every fixed $1<c<\sqrt2$, we maintain an actual factor-$c$ diameter pair under arbitrary point insertions and deletions, with worst-case update time $n^{2/c^2-1}\poly(d,\log n)$ and space $n^{2/c^2}\poly(d,\log n)$, suppressing logarithmic dependence on the execution length and failure probability. Here $n$ bounds the number of active points. The update exponent improves Indyk's algorithm~\cite{indyk} (SODA'03) bound from $1/c^2$ to $2/c^2-1$, while the space exponent improves from $1+1/c^2$ to $2/c^2$. At $c=\sqrt{3/2}$, the update exponent decreases from $2/3$ to $1/3$. A separate construction achieves approximation exactly $\sqrt2$ with update time polynomial in $d$ and $\log n$. All these guarantees hold against adaptive updates, even when every previous random choice is exposed. We also improve the time--space tradeoff near linear space. This construction has an always-correct variant with expected amortized update bounds and requires no advance bound on the number of updates. The dynamic structures store a current answer, so reporting its two point identifiers takes constant time. The geometric method also gives static diameter algorithms and a farthest-neighbor structure for a fixed set that answers arbitrary adaptive queries. At approximation $\sqrt2-\varepsilon$, the diameter time is $n^{1+O(\varepsilon)}$, and the query time is $n^{O(\varepsilon)+o(1)}$ with near-linear preprocessing and space, up to polynomial input-parameter factors. On binary points, this improves the $n^{O(\sqrtθ)}$ overhead of the algorithm that Andoni, Jiang, and Zharkov~\cite{ajz2026} (STOC'26) for Hamming approximation $2-θ$ to $n^{O(θ)+o(1)}$.

Most AI assistants respond only when prompted, remaining silent even when events in the video stream offer meaningful opportunities for conversation. To move beyond purely user-driven interaction, we present AURA 2.0, a streaming video large language model that introduces three core upgrades to AURA. The key advance is Model-Initiated Conversation (MIC), which enables the model to identify topics worth discussing in the evolving visual stream and initiate conversations at appropriate moments without a user query. Stream-Aware Tool Calling (SATC) enables the model to invoke external tools using the current visual and conversational context and incorporate returned information into the ongoing interaction. Interactive Visual Grounding (IVG) supports visual grounding in both user inputs and model outputs. The model can identify objects marked by users on video frames and produce frame-aligned bounding boxes to link its responses to specific objects and regions. To develop these capabilities, we construct a dedicated streaming interaction dataset and adopt a three-stage training scheme to integrate the new behaviors with AURA's existing streaming interaction abilities. We further incorporate Long-Horizon Interaction Memory (LHIM) to maintain contextual continuity over extended video streams. Evaluations on public streaming benchmarks show that AURA 2.0 achieves leading performance among open-source streaming models, while targeted evaluations further validate its new interaction capabilities. We open-source the model weights and inference code.

Perceptual quality enhancement of severely compressed videos remains challenging due to complex artifact patterns and substantial information loss. Recent diffusion models have demonstrated strong generative capability for visual restoration, but directly applying them to compressed video often ignores compression degradation characteristics and may introduce structure-inconsistent hallucinations. To address this issue, this paper presents a diffusion-based compressed video enhancement method, named DiffCVE. Coding Prior-enhanced Dual Conditioning (CPDC) branches are designed to jointly model compressed video and coding prior conditions, where coding priors including residuals and motion vectors provide complementary structural and motion guidance during the diffusion denoising process. To make the diffusion process aware of compression severity, a Compression Degradation Semantic Prompting (CDSP) mechanism is introduced to leverage QP-conditioned textual prompts together with LoRA fine-tuning. In addition, a Coding Prior-guided Weighted Fusion (CPWF) module is incorporated into the VAE decoder to fuse VAE encoder and coding prior encoder features with QP-predicted weights. Extensive experiments demonstrate the effectiveness of the proposed method in improving perceptual quality, especially under severe compression settings. The project page and code are available at https://wqmaker.github.io/projects/DiffCVE/.

Large Language Models (LLMs) are increasingly used for database tasks, yet most work, such as Text2SQL, assumes that the relational schema already exists. We ask whether LLMs can design a schema from natural language requirements (Text2Schema), verbalize an existing schema (Schema2Text), and preserve its structure through a Schema-to-Text-to-Schema round trip. We present SchemaBridge, a benchmark of 902 schemata from three collections, stratified by size and constraint complexity, that scores fidelity separately for tables, attributes, domain (data types and nullability), entity, and referential integrity. Evaluating six open-weight and proprietary LLMs, we find that: (i) human-written requirements under-specify schemata, with no model exceeding 0.27 referential integrity F1; (ii) LLM-generated descriptions carry far more structure when explicitly asked to be complete, raising GPT-5.4-mini's referential F1 on SchemaBench from 0.228 to 0.909, even as BLEU, ROUGE-L, and BERTScore decrease; (iii) controlled experiments show that both constructor and describer choices influence round-trip fidelity under fixed settings; and (iv) masked schema experiments show that structural cues support recovery of keys and relationships, while joint data-type and nullability information is poorly retained through verbalization. Code and data are available at https://github.com/iamsriom/SchemaBridge

Let $m,n,d >1$ be integers such that $n=md$ and let $(B, \mathscr{M}, Θ\_{\mathscr{M}})$ be a marked abelian variety of level $m$ over a field of odd characteristic. We study the theta structures of level $n$ for $\mathscr{M}^d$, and the theta structures on the abelian varieties $d$-isogenous to $B$, which are compatible with $Θ\_{\mathscr{M}}$. When $d$ is odd and coprime to $m$, they correspond to the isotropic subgroups of $B[n]$, and this correspondence underlies the known algorithms in theta coordinates. In general, this is no longer true. We introduce the notion of symmetric compatibility of a torsion point with a symmetric level subgroup, and we show that a subgroup of $B[n]$ isomorphic to $(\mathbb{Z}/n\mathbb{Z})^g$ comes from a compatible theta structure if and only if it is isotropic and all its points are symmetric compatible. We extend to every $d$ Mumford's notion of compatible pairs of theta structures, defined by Mumford for $d=2$, and prove that it is equivalent to our notion, which can be tested on theta coordinates; we also describe exactly the remaining ambiguity, which explains the sign choices met when the duplication formula is used to change the level. As applications, we obtain change of level and isogeny algorithms without assuming that $d$ is odd and coprime to $m$, using $O(n^{3g} \log(d))$ and $O(n^{2g} \log(d))$ operations in the base field respectively. In particular, they give change of level algorithms for even $d$, and $2$-isogenies in level $4$ theta coordinates with an explicit control of the theta structure of the codomain.

Users of artificial intelligence services, including those powered by large language models and other transformer-based models, need assurance that inference and training follow the declared computation without exposing private models or data. Zero-knowledge proofs can provide this assurance, but proof generation remains expensive. Hardware acceleration research often reports gains for individual cryptographic operations, leaving their effect on the full cost of proving neural-network workloads, particularly large transformer models, unclear. We introduce a four-layer framework linking verifiable AI workloads, proof systems, cryptographic kernels and hardware platforms, use it to review the literature, and add original measurements of six provers and seven kernels to identify which costs matter in practice. Our measurements show that the dominant cost changes with the proof system and input length. In low-rank fine-tuning, the cost of opening commitments to fixed model weights remains nearly constant across input lengths from 64 to 2048, while the work associated with intermediate values grows. A 13-billion-parameter fine-tuning configuration exceeds the capacity of a 40-gigabyte graphics processor but completes on an 80-gigabyte device, showing why capacity must be assessed alongside execution time. Our kernel profiling further identifies instruction issue and synchronization limits that a bandwidth-only model misses. We use these findings to assess graphics processors, reconfigurable logic, custom chips and existing artificial intelligence accelerators. The resulting research agenda connects proof-friendly numeric formats, reuse of fixed-weight commitments, heterogeneous execution and shared benchmarks to the measured bottlenecks. Effective acceleration requires selecting hardware and proof-system optimizations around the cost profile of the target workload.

Multi-agent deep-research systems use delegation to parallelize search and keep long tool traces out of the main agent's context. Yet delegation helps only when task allocation and subtask execution are coordinated: the main agent must issue useful subtasks, while subagents must return reliable evidence. Existing training either optimizes only the main agent or broadcasts the same team advantage to all agents, which can reward misleading subagent reports and penalize useful ones. We introduce LEAD-RL (Learning Effective Agent Delegation with Reinforcement Learning), a joint main-agent-subagent training framework with contribution-aware credit assignment. A training-time rubric evaluates the research utility of each subagent report and gates its inherited team advantage, while final-answer correctness remains the only team reward; no token, step, or speedup reward is used. Across five benchmarks and two model families, LEAD-RL improves answer quality. On BrowseComp, it improves accuracy by 2.49 percentage points and reduces critical-path cost by 24.5% over joint training with shared team credit. During validation, accuracy increases by 6.78 points at approximately constant critical-path cost. These results show that contribution-aware credit assignment improves the quality-efficiency trade-off of multi-agent deep research.

Watermark detectors decide whether one document came from a watermarked language model. Auditors often need a different number: the share of a corpus that did. We study this share for Gumbel-max and green-list (KGW) watermarks when every document may have been edited by an unknown amount. Without restrictions on the edits, the upper end of the identified set is 1. Combining a known least-favourable next-token distribution with a sequential coupling, we derive a power envelope indexed by three budgets: a retention floor, an entropy budget, and a budget for in-place substitutions, which for Gumbel-max make edited positions look less watermarked than human text. The envelope gives an identified set that is sharp for the alternatives it dominates, finite-sample bounds on the realized share of a fixed corpus with either the exact null or a null calibrated on an independent human sample, and confidence sequences for streams. In Gumbel-max simulations where the envelope binds, the bounds cover, and the threshold that maximizes the gap between envelope and null is generally as tight as twelve thresholds. On text from OPT-1.3B and Qwen3-1.7B, 12,000 fresh human documents scored under three keys show no robust departure from the exact null, and all bounds covered in resampling trials from finite pools of documents with random edits, substitutions and paraphrases; at a share of 0.2 the exact-null bound is about [0.14,0.38]. These results are fragile in two ways we make precise. LLM-paraphrased KGW pools fall below the envelope, yet at 2000 documents the bounds still covered in these trials. A simulated null error of 0.01, comparable to the largest deviation we observed at our thresholds, is likewise harmless at 2000 documents but breaks exact-null coverage at 100,000; a null calibrated on an independent human sample keeps coverage there, but needs a large reference sample to stay narrow.

Electroencephalography (EEG) foundation models are pretrained on thousands of hours of recordings, yet their frozen features transfer poorly to downstream tasks, and this transfer does not consistently improve with larger models or more pretraining data. We show that a vision foundation model that has never seen EEG is a stronger frozen EEG encoder than these specialized models, once the signals are rendered as images that a vision transformer can read. Our method, SeeEEG, keeps the vision backbone frozen and learns only how EEG is displayed to it. Each channel occupies a dedicated row of patches and is drawn either as a waveform at a fixed gain or as a scalogram whose frequency range spans the patch height, so that task-relevant patterns fall within individual tokens. The montage, filtering, and contrast parameters of this display are initialized from a standard clinical montage and optimized with the task loss through the frozen backbone, which amounts to at most 1,113 trainable parameters per view. We compare SeeEEG with seven EEG foundation models on eight clinical, sleep, affective, and brain-computer interface datasets, with every encoder frozen and paired with identical readout heads. With a frozen DINOv3 ViT-B, SeeEEG outperforms every EEG foundation model in balanced accuracy on all eight datasets and reaches a mean rank of 1.3 in Cohen kappa or AUROC, compared with 2.9 for the strongest baseline, REVE. The learned displays transfer without further optimization to the video encoder V-JEPA 2.1, which outperforms every EEG foundation model on both metrics on six of the eight datasets. Replacing DINOv3 with a randomly initialized ViT-B lowers the score by 15.2 points on average, so the ability to read rendered EEG comes from visual pretraining.