Papers

The explosive growth of short videos has presented significant challenges to content governance. Traditional manual labeling is inefficient, slow to iterate, and costly. Therefore, we explore and develop a multi-stage autopilot labeling framework aimed at progressively increasing the level of automation in short video content annotation. To move beyond the heavy reliance on human expert labeling at the initial L0 stage, we first introduce a discriminative MLLM labeler at the L1 stage, which is fine-tuned with supervised data to directly predict labels for short videos. Advancing to the L2 stage, we shift to a domain-adaptive MLLM pretrained on short video data and use it for zero-shot labeling, enabling flexible generalization without task-specific supervision. In the L3 stage, we propose an agentic annotation workflow, and equip the agent with several key tools, which empower the agent to handle complex, unfamiliar, or vaguely defined cases, significantly reducing the need for human intervention. We conduct extensive experiments across various content issues, and results demonstrate that each stage of the autopilot labeling framework progressively improves labeling accuracy, and has significantly reduced manual labor costs in real-world applications by up to 85%.

Conventional distillation transfers capability through teacher-generated demonstrations. LLM-as-a-Judge knowledge distillation offers a distinct and increasingly practical alternative. In the label-only variant studied here, the student generates every response and improves from scalar scores alone. Prior work shows that this non-demonstrative feedback can improve student capability, but it creates a fundamentally different provenance problem. No teacher-generated answer enters the student's training set, so any evaluator-specific trace must pass through a low-bandwidth reward signal rather than copied text. Existing watermarks and distillation detectors do not cover this score-only channel. We introduce Radioactive Feedback, which encodes a registered pattern of behavioral preferences in the rewards themselves. The evaluator applies bounded pointwise adjustments to quality-qualified responses, allowing repeated optimization to transfer the pattern to a student trained by another party. Experiments on code and mathematics show that the registered preferences transfer through scores alone and across datasets, enabling black-box detection against a fixed reference pool.

A key milestone in algorithmic graph theory is Hopcroft and Tarjan's 1974 linear-time algorithm for graph planarity testing. At a workshop in 2002, Linial and Rabinovich asked for either a polynomial-time algorithm or an NP-hardness proof for metric planarity testing: given a metric space $(X,d_X)$, determine whether there exists an edge-weighted planar graph $G=(V,E,w)$ with $X\subseteq V$ that realizes the distances in $X$, meaning that $d_G(x,y)=d_X(x,y)$ for every $x,y\in X$. This question arose in the context of research on metric embeddings and their applications in computer science since the mid-1990s. Over the past few decades, polynomial-time algorithms have been developed for several special cases. In this work, we show that metric planarity testing is NP-hard, in sharp contrast to graph planarity testing. Our reduction yields a stronger result: under the Exponential Time Hypothesis (ETH), no $2^{o(n)}$-time algorithm for metric planarity testing exists, where $n$ is the number of points in the metric. To complement our ETH lower bound, we give two algorithmic results: (i) an algorithm for computing a metric planar embedding in $2^{\tilde{O}(n^2)}$ time and (ii) a $(1+\epsilon)$-approximation algorithm in $2^{\tilde{O}(n)}$ time for any fixed $\epsilon\in(0,1)$. The second algorithm produces a planar graph that preserves the input metric up to a factor of $(1+\epsilon)$.

Parametric Computer-Aided-Design (CAD) modeling from human intent remains challenging, particularly during the conceptual design stage, where design goals are expressed through incomplete and unstructured modalities (e.g., hand-drawn sketches and textual descriptions). In this work, we rethink the human intent-to-CAD pipeline and propose a unified method that directly maps multi-level human intents to executable codes, without assuming the prior existence of target CAD models. To support our study, we construct HiCAD, the first large-scale dataset aligning hand-drawn sketches, textual descriptions, and parametric CAD codes. Based on this, we introduce HiCAD, a two-stage framework comprising Cooperative Multi-Task Alignment to bridge the representational gap between heterogeneous inputs, and Spatial-Aware Reinforcement Learning to enforce geometric and topological consistency. Extensive experiments demonstrate that our method significantly outperforms existing baselines across multiple tasks, validating its effectiveness and robustness in transforming heterogeneous human intents into high-fidelity parametric CAD models. Our project page: https://zqwlearning.github.io/HiCAD.

Carl Jung theorised that the human psyche contains an inherited framework, the collective unconscious, populated by universal archetypes. As large language models (LLMs) are trained on vast corpora of textual data, this paper investigates whether they spontaneously construct distinct archetype-like regions within their high-dimensional latent spaces. The paper proposes that these regions function as spatial constraints that naturally organise semantically related narratives. To evaluate this, we applied cosine similarity and angular distance metrics across thirteen experiments utilising OpenAIs text-embedding-3-small model. The paper contains two experimental sections the first section explores these ideas using synthetically constructed narratives, the results of which act as a baselines for experiments conducted in the second section where the synthetic data is replaced by human-authored literature (Shmoop corpus), unedited dream reports (DreamBank corpus), and culturally distilled folklore (Andrew Langs Fairy Books). The results show that archetypal roles occupy distinct, measurable regions in the latent space. Crucially, these regions remain compact and consistent because of narrative structure rather than simple word overlap. Additionally, text category dictates spatial distribution: dream reports concentrate within Trickster and Shadow regions, while traditional folklore distributes evenly across all tested archetypes. We conclude that transformer latent spaces encode stable regions which appear to correspond to recognised Jungian archetypes.

Continual learning aims to adapt pretrained large language models (LLMs) to evolving data while preserving their pretrained capabilities. A central challenge is catastrophic forgetting: adaptation to new data can impair performance on previously learned knowledge. Since continual learning commonly changes model parameters, the model's predictions may draw on knowledge from the base model, the parameter update, and the input context. Understanding how these sources interact is essential: conflicts reveal whether new knowledge can override outdated knowledge, while multi-hop reasoning tests whether new knowledge can be composed with existing knowledge to support generalization. We study these interactions in a hypernetwork-based adaptation framework, where a shared network generates adapters from context, allowing a frozen base model to adapt without task-specific updates. We find that, under knowledge conflict, models preferentially follow the context, and their choices depend more on where knowledge is presented than on its content. We also show that context and adapter knowledge perturb the base model's representations along directions orthogonal to its clean hidden states, enabling us to derive steering vectors that control which source the model follows. In multi-hop reasoning, we find that models fail to reliably compose knowledge across the three sources, with the key breakdown arising from the interaction of the hypernetwork with both the base model and the context. Code and reproduction materials: https://github.com/missswiftie/Knowledge-Interaction-in-CL

In this paper, we develop a greedy double-block nonlinear Bregman-Kaczmarz method to efficiently find sparse solutions of nonlinear equations. The theoretical convergence rate of the proposed method is established under certain assumptions. Numerical experiments illustrate the competitive performance and superiority of the proposed method.

Quantization lowers deployment costs of vision-language-action (VLA) policies but can compromise task completion. Across five VLA models and 4,050 real-world trials, we systematically reveal that quantization-aware training (QAT) generally improves on post-training quantization (PTQ) but leaves substantial gaps to full precision (FP) on long-horizon tasks and under visual and semantic shifts, while its imitation-based objective does not explicitly use rollout success or partial completion. To exploit these outcome signals, we introduce Quantization-Conditioned Reinforcement Learning (QCRL), an offline RL method coupling progress-based advantage estimation with quantization-aware source conditioning. FP rollouts supply reference experience beyond action-level distillation, progress-scaled terminal penalty distinguishes partial completion, and quantization-aware conditioning selects advantageous reference behavior within the quantized policy. We evaluate weight-only INT4 W4A16 on $\pi_{0.5}$ and GR00T N1.5 and joint weight-activation NVFP4 W4A4 on $\pi_{0.5}$. Using only in-distribution data, QCRL reaches 99.3% of FP's mean score across three visual-shift conditions for $\pi_{0.5}$ with INT4 W4A16, alongside 42.9% model-memory savings. With NVFP4 W4A4, it reaches 100.7% of the same FP baseline, alongside a $1.73\times$ speedup using NVFP4/FP8 deployment. Hardware gains are relative to FP16 on Jetson Thor.

Eyewear that senses articulation with inaudible ultrasound can now recognize silent speech over an open vocabulary, but its accuracy is limited by the cost of training data. We present SoniGAN, which synthesizes the ultrasonic echo profiles of silent articulation from text and uses them as synthetic supervision. A voice-cloned text-to-speech (TTS) model reads a text corpus, and a two-stage generator synthesizes echo profiles from representations of the spoken content of that audio. Pretraining a recognizer on 104 hours of the synthesized data improves the word error rate (WER) of a public open-vocabulary benchmark by 2.3 points. Adding the sessions the corpus discards for a faulty microphone brings the system to 22.2% without any language model. Lexicon-constrained decoding, LLM rescoring, and ensembling further bring it to 14.9%, the first minimally obtrusive silent speech interface to fall inside the error range in which users report no loss of willingness to use silent speech input.

Weak bag-level labels are widely available in computational pathology, whereas reliable instance-level annotations for whole-slide images remain expensive and ambiguous. Weakly supervised multiple instance learning (MIL) achieves strong bag-level classification, but instance-level supervision from attention scores, model predictions, negative bags, or spatial refinement may fail to identify class-specific local evidence. We propose ProtoPur, a prototype-purity pseudo-labeling method for weakly supervised instance-level learning in pathology whole-slide images. ProtoPur estimates class-balanced local purity over prototype regions constructed from frozen foundation-model features, and assigns pseudo-labels to high-purity regions strongly associated with a target class. This purity-first design separates class-specific evidence from negative or shared morphology, with coverage used only to ensure sufficient pseudo-label assignment. We provide a PAC-style analysis showing that the risk of an instance classifier trained with ProtoPur pseudo-labels can be decomposed into a standard uniform-convergence term and a ProtoPur-specific pseudo-label noise term, comprising selected-region mismatch and a coverage residual. Experiments on three public whole-slide image datasets with two pathology foundation encoders show that ProtoPur consistently improves instance-level performance across multiple MIL backbones while maintaining competitive bag-level performance. Pseudo-label quality comparisons on two public datasets further demonstrate that prototype-purity selection provides stronger instance supervision. These results suggest that the foundation-model features contain exploitable instance-level supervision signals, and ProtoPur is a way to reveal and use these signals from weak bag labels without requiring dense instance annotations.

A single evaluation score can rank generation systems but rarely reveals what failed. We propose rate-resolved text evaluation, which turns the masking rate of masked diffusion language models into a diagnostic axis. By reconstructing text under different fractions of masked tokens, one diffusion score becomes a vector of band-conditioned scores that reflect reconstruction difficulty at varying levels of visible context. On SummEval, preserving this rate structure yields complementary signal: the best masking-rate band improves fusion with autoregressive likelihood over a collapsed diffusion scalar, survives a sample-shuffle control that breaks rate ordering, and adds residual signal to a fixed prompted judge. A second diffusion model reproduces the effect more weakly. On WMT MQM zh-en, roundtrip mask-and-reconstruct probing produces a tunable localization curve in which ranking-oriented AUROC and span-detection F1 peak at different masking rates. Factuality benchmarks bound the interpretation: band semantics are model- and task-dependent, so each application requires empirical validation.

Converting natural-language instructions into executable action plans conditioned on a scene image is a core capability for embodied human-robot interaction, but existing benchmarks target English, simulated or video-grounded environments, and largely assume instructions are clear and executable. We introduce the first Persian vision-language-action dataset for this task: 258 instances over 230 real household/office photographs, annotated with reference plans, deliberate ambiguity and infeasibility labels, disambiguating question-answer pairs, and an auto-generated chain-of-thought rationale for training. Using this resource, we ablate retrieval-augmentation, chain-of-thought prompting, and an agentic disambiguation step, evaluated with an LLM-as-judge protocol and a new metric, PGOS-FA, extending prior work (PG2S) with step-order sensitivity, an over-generation penalty, and Persian-adapted embeddings. Paired statistical testing across 36 comparisons shows retrieval-augmentation is a large, robustly significant effect, while finer-grained additions are directionally consistent but not statistically distinguishable from noise at this scale, and the LLM-judge metric has near-zero power where PGOS-FA succeeds. A blind human-correlation study on 79 real outputs shows PGOS-FA tracks human judgment at least as closely as the LLM-judge. Testing generalization beyond the ablation's single backbone, we find that most of the gap between an open-weight and a substantially stronger closed planner is attributable to open-weight model generation rather than to the closed model's capability specifically, and that our best configuration significantly outperforms an independently reproduced external baseline despite a weaker backbone. The dataset, PGOS-FA implementation, and evaluation code will be made available upon reasonable request.

We present AnyEdit, a unified, separation-free framework for speech and singing voice editing that preserves real-world acoustic environments, including noise, reverberation, and background music. Recent unified generative models already achieve strong speech and singing generation in clean conditions, but real-world vocal editing remains challenging because the model must revise content while simultaneously preserving speaker identity, prosody, and the surrounding acoustic scene. Directly modeling noisy audio end-to-end entangles semantic editing with environmental acoustics, weakening the core vocal generation capability. We therefore adopt a two-stage design that preserves clean-domain editing ability while deferring acoustic rendering to a separate stage: an instruction-guided autoregressive editor first predicts clean prosody and content-style tokens, supported by a teacher-student noise-robust prosody tokenizer that distills clean prosodic tokens from degraded recordings; a flow-matching acoustic model then performs in-context rendering conditioned on the source acoustic context, so the edited segment inherits surrounding noise, reverberation, and background sound while maintaining consistent speaker timbre. To stabilize unedited regions, we introduce unmodified-area-aware supervision; after pre-training, we further apply direct preference optimization on mined real-to-real singing pairs. We also release AnyEditBench, a benchmark covering both speech and singing voice editing under clean, simulated, and in-the-wild acoustic conditions. Experiments on Chinese and English speech and singing datasets showed that AnyEdit consistently outperforms strong baselines in content accuracy, subjective quality, and background consistency under noisy and reverberant conditions. Demo: https://any-edit.github.io/. AnyEditBench: https://huggingface.co/datasets/Ririiiii/anyeditbench.

We give a simplified proof of the polynomial improvements for integer 3SUM and all-pairs shortest paths (APSP) established by Alman and Vassilevska Williams, with weaker exponents. The central ingredient is selected-entry matrix multiplication: computing a prescribed set of entries of a product of dense rectangular matrices. Starting from Schönhage's inner/outer-product identity, we construct a multiplication rule that shares intermediate products among many outputs. We apply this rule recursively and bound the work by counting the requested entries at each level. Applied to incidence matrices, the algorithm solves Offline Set Disjointness over a small universe. Known deterministic reductions then yield truly subquadratic integer 3SUM and truly subcubic APSP for polynomially bounded integer inputs.

Generative policies offer expressive action distributions for online reinforcement learning (RL), but many diffusion- and flow-based policies require costly iterative sampling. We propose DriftSpread, a one-step generative policy based on Drifting for online maximum-entropy (MaxEnt) RL. To incorporate entropy without a tractable policy density, we build on an energy-based view of Drifting and derive critic-guided Drift and entropy-induced Spread from the negative MaxEnt objective. Our method learns these updates sequentially in the same actor, using action regression for value improvement and distribution matching to capture the effect of entropy. The actor updates require neither direct samples from the critic-defined Boltzmann target nor policy-density or score estimation. Under exact fitting with a fixed critic, we prove convergence to the Boltzmann target in the small-step limit followed by the long-time limit. On MuJoCo, DriftSpread achieves the highest aggregate normalized return among all evaluated methods, while outperforming the diffusion- and flow-based baselines in both training and inference efficiency.

Feedback presentation shapes how LLMs use evaluator judgments during translation refinement. We introduce a machine translation protocol that keeps the evaluator score, main issue, and recommendation fixed while varying feedback form and supporting detail. Across seven models and six translation directions, low-load feedback outperforms high-load feedback in all 42 translator-direction settings under xCOMET. On the 29,400-record rewriting subset, mean xCOMET scores multiplied by 100 are 83.30 for low-load feedback and 80.67 for high-load feedback; the full-matrix original-translation reference mean is 83.46. MetricX-24 supports the low-load advantage on a stratified subset. A no-feedback second pass yields a small improvement over the original translations, and direct evaluator prompting reproduces the low-load advantage without a separate rewriting stage. These results show that compact feedback limits the quality loss from evaluator-guided refinement in the evaluated high-resource MT settings.

Domain generalised semantic segmentation (DGSS) has recently benefited from combining multiple foundation models, such as vision-language and vision foundation models, to leverage their diverse representations for generalisation. Existing multi-model approaches share a common design choice: cross-model fusion is restricted to depth-aligned layers, implicitly assuming that mutually beneficial information resides at the same depth across models. However, foundation models pretrained under different objectives may develop distinct representation hierarchies, making the optimality of depth-aligned fusion questionable. Moreover, we observe that the foundation models exhibit different depth-wise domain sensitivity. Motivated by these observations, we propose a Sparse All-Layer Connector (SALC) that enables each layer to access information from preceding layers of another model. As the number of accessible layers grows with depth, dense aggregation may hurt generalisation. SALC therefore learns to adaptively select a sparse subset of informative layers. We further introduce a candidate dropout regularisation that strengthens sparsity and encourages SALC to explore diverse layers, leading to more robust selection. Across four foundation model combinations and three DGSS evaluation settings, the proposed design demonstrates stronger generalisation than conventional depth-aligned fusion. The code is available at https://github.com/RuoyuGuo/SALC.

Centrality measures rank nodes even when the observed differences can be explained by ordinary structural variation. We therefore study a preceding question: whether a network contains a statistically identifiable central structure at all. We formulate this question as a network-level hypothesis testing issue using the adjacency spectral radius as the test statistic and estimate an empirical p-value from randomized networks generated under a specified null model. The framework is evaluated using G(N,M), Chung-Lu, and two implementations of the configuration model on simulated benchmarks and 55 real networks. Controlled experiments further vary the number and connection weights of designated central nodes while holding network size and density fixed. The results show a clear negative association between central-structure strength and the empirical p-value. They also demonstrate that rejection decisions depend substantially on the structural properties preserved by the null model: the detected network sets are not simply nested across models.

Distributed Constraint Satisfaction Problems (DCSPs) model distributed decision-making in multi-agent systems, where agents assign values to local variables while satisfying shared constraints. However, many DCSP solvers still rely on broad, repeated peer-to-peer communication and are largely reactive, with limited support for strategic coordination or learning. This increases communication overhead, can lead to stagnation, and reduces the success of hard instances. Hence, we propose Dynamic Conflict Clustering-Incomplete (DCC-Incomplete), a repair-based solver that combines conflict-centred decomposition, structured coordination, strategic repair, and lightweight learning for stagnation recovery. DCC-Incomplete groups the unsolved part of the problem into conflict clusters and restricts communication via a coordination layer to agents within the same conflict region. It also introduces memory-guided restart, enabling agents and clusters to pause, revise, and adapt repairs based on neighbouring conflict states. A thorough evaluation of DCC-Incomplete on distributed graph-colouring benchmarks with planted satisfying assignments shows that its memory-guided restart improves the success rate from 37.50% to 61.11%, reduces median runtime from 0.565s to 0.367s, and cuts median communication from 12.493 to 6.503 messages per constraint. These results suggest that dynamic conflict clustering, combined with lightweight strategic and learning-guided repair, improves both communication efficiency and robustness in incomplete DCSP solving.

We study the communication cost of exact sampling under Campbell's exponential coding criterion. Given a target distribution $P$ and a proposal distribution $Q$, the sender and receiver share i.i.d. $Q$-distributed samples, and the sender communicates an index so that the selected sample has distribution exactly $P$. We introduce a new information measure Rényi channel simulation divergence $R_\alpha(P\Vert Q)$, a Rényi generalization of the channel simulation divergence. We then show that it lower bounds the Campbell cost of every exact sampler and is achievable, via a new layered sampling scheme, within an additive constant depending only on the Campbell parameter. This yields a tighter one-shot characterization than previous bounds based on Rényi divergences of different orders. For i.i.d. product distributions, under suitable non-degeneracy and moment conditions, we further show that the gap beyond the leading Rényi-divergence term is $\frac{1}{2}\log n+O(1)$.

Hive is a boardless two-player strategy game in which the pieces form the playing surface. We prove that deciding whether White can force a win in single-Queen $n$-Hive, the generalization with exactly one Queen Bee per player, is EXPTIME-complete, even when all pieces are already placed. Our hardness proof reduces from unbounded two-player Constraint Logic: Grasshoppers encode edge orientations, while the One Hive rule enforces the vertex constraints. We obtain the EXPTIME upper bound by considering positions up to translation. The result answers the single-Queen question raised by Andel and Rin. Keeping the same inventory and movement rules, it also gives EXPTIME-completeness for multi-Queen variants. To the best of our knowledge, this is the first use of unbounded two-player Constraint Logic to prove EXPTIME-completeness for a natural problem.

Fourier neural operators predict power-system transient trajectories with high accuracy. We extend this line of work by examining the choice of spatial spectral basis, studying both the graph Fourier transform derived from the grid's own coupling network and learnable bases trained from that starting point and from a Fourier control. Evaluated in one controlled pipeline across three protocols spanning unseen fault locations and shifted operating regimes, the learnable bases improve on a parameter- and budget-matched DFT variant by 7-14% in distribution, with per-case significance on both in-distribution protocols and no loss on the third; the gain comes from adaptation rather than from the coordinates alone. We further characterize the regime dependence of this gain, identifying its cause, locus, and continuity, and show that a resolution requires no new model: averaging the learned and frozen variants gives the best shift-robust predictor, while their disagreement provides a label-free alarm that flags every shifted case with no false alarms.

High-order direct arbitrary Lagrangian–Eulerian discontinuous Galerkin (ALE-DG) methods can suffer severe order reduction under nonsmooth mesh motion despite satisfying the geometric conservation law (GCL). To address this limitation, we propose a new ALE-DG framework applicable to moving triangular and quadrilateral meshes. This framework successfully overcomes the uniform Lipschitz constraint inherent in existing direct ALE-DG methods, retaining high-order accuracy even when the mesh-velocity gradients are unbounded. This capability is driven by transport polynomial exactness (TPE($k$)), which is a structural property stronger than the GCL. TPE($k$) ensures the exact transport of polynomials of degree at most $k$ under arbitrary mesh motion. To incorporate TPE($k$) into a direct ALE-DG method, we introduce a coefficient-evolution (CE) formulation. Reconciling strict local conservation, an indispensable requirement for accurate shock capturing, with the preservation of the TPE(\(k\)) property constitutes the principal technical challenge. We resolve this by evolving the coefficients using the third-order strong-stability-preserving Runge–Kutta method, paired with a conservative correction executed via Simpson's rule integration. Crucially, this correction strictly vanishes for exactly transported quadratic polynomials, thereby ensuring that our third-order scheme rigorously satisfies the TPE($2$) property. Numerical experiments demonstrate that the proposed scheme retains third-order accuracy when the discrete mesh-velocity Lipschitz estimate grows as $\mathcal O(h^{-1})$, whereas a representative existing direct ALE-DG scheme exhibits order reduction under the same mesh motion. Tests with discontinuous solutions further demonstrate robust and high-resolution performance under random and Lagrangian mesh motions.

This article aims to design a distributed economic dispatch (ED) scheme based on marginal cost (MC) consensus for a battery energy storage system (BESS) network considering the utility grid switching. Due to the low control accuracy and slow convergence speed of previous schemes, some important changes have been made in this article. Firstly, the feedback coupling term is introduced into the closed-loop error to further improve the stability and convergence rate of the controller. Then, a proportional integral (PI) controller with a reset mechanism is implemented in the designed scheme, which improves the speed and control accuracy of MC consensus. The performance of a control system with a reset mechanism is well guaranteed by analyzing the eigenvalues of the corresponding matrix. Finally, some tests focus on UG switching, time phased pricing, load switching, a BESS/agent plug and play, and large-scale system applications to illustrate the effectiveness and progressiveness of the design scheme.

Surgical world models must turn prospective robot actions into visually faithful, structurally consistent instrument motion. We introduce SurgCast, which predicts future instrument skeletons from current structure, robot state, and future actions, then uses the forecast to control video generation. Dual-path injection combines skeleton geometry with derived, instrument-specific motion descriptors through geometric residuals and conditioning modulation. Both paths complement native action conditioning and are retained in a four-step causal student. On SutureBot and SRTH-Porcine-Cholei, the teacher and student improve all seven reported quality and instrument-consistency metrics over their respective direct-action baselines; teacher FVD decreases by 37.6% and 67.3%. Ablations show that future actions improve skeleton prediction and that combining both injection paths performs best across all six metrics under predicted skeletons. These results support using predicted articulated structure and its motion semantics as a control interface for surgical video prediction. Project videos and examples are available at https://surgcast.netlify.app/.

Under a limited context budget, a long-term agent must recall not merely memories that are individually relevant, but evidence that is collectively sufficient for the current query. Conventional memory access retrieves, scores, and truncates memories, often spending multiple context slots on redundant facts while leaving complementary evidence behind. We cast the read phase of long-term memory as query-conditioned evidence composition: each memory is valued by what it adds beyond the evidence already selected. We instantiate this operation with MemPursuit, a training-free mechanism that uses the residual of a sparse convex query reconstruction as a representation-space proxy for what the current memory set is missing. By reusing the query and candidate embeddings already available from retrieval, MemPursuit progressively composes nested evidence contexts without invoking a language model or cross-encoder to score each query–memory pair. The effectiveness of our method is verified on 3 popular benchmarks in comparison to strong baselines. In fixed-budget LoCoMo experiments, MemPursuit improves end-to-end accuracy by 7.2 percentage points at Top-10 and 5.5 percentage points at Top-20, with consistent gains across multiple memory frameworks, while achieving a 32.9$\times$ speedup in median reranking latency over fully batched BGE-v2-m3. Our implementation is available at https://github.com/FengQi-HITSZ/MemPursuit.

This paper investigates how bistatic delay information improves the observability and localization performance of 3D moving targets in distributed Wi-Fi sensing. Motivated by emerging IEEE 802.11bf WLAN sensing, we consider one transmitter, multiple distributed receivers, and a single target with unknown position and velocity. We compare Doppler-only sensing with delay-Doppler sensing, in which delay-capable receivers provide additional bistatic path-length information for target localization. We derive the Fisher information matrix (FIM) and the position equivalent FIM (EFIM) for both sensing formulations. The resulting observability analysis shows that Doppler-only localization generically requires six receivers, whereas delay-Doppler localization can become locally identifiable with three receivers under non-degenerate geometry. We further study mixed receiver architectures in which only a subset of receivers provides delay estimates. Simulations with Wi-Fi-like bandwidths confirm the transition and quantify the roles of bandwidth, Doppler noise, receiver geometry, and delay-capable receiver availability.

Small language models (SLMs) used as on-device agents must balance accuracy and latency as their context grows. Mixed precision helps manage this tradeoff by assigning different precisions to weight tensors according to their size and sensitivity. However, these choices depend on the operating conditions, so a configuration selected for a short context may no longer provide the best tradeoff at a longer one. We calibrate a timing model from measurements at different precisions and context lengths, and combine it with tensor sensitivities to choose configurations under a latency or accuracy budget. We argue that fewer bits do not always mean faster execution, so our predictor estimates latency from calibrated execution costs rather than weight bytes alone. Under a latency budget, reading the larger key-value (KV) cache leaves less time for weight operations, leading the allocator to select lower precisions. This gives a schedule of tensor changes that can be prepared before execution. Our runtime follows the schedule while retaining the session's KV cache. We evaluate executed schedules on a laptop GPU, including a longer-context schedule with three weight formats and a tool-use benchmark measuring full response time and correctness. Under an accuracy budget, measured error from a quantized KV cache determines the budget left for weights. The allocator can raise or lower weight precision as this estimate changes. These experiments show where context-aware precision helps and where accuracy recovery remains limited.

Long-horizon robotic manipulation is inherently composed of semantically meaningful sub-tasks. While Movement Primitives (MPs) can model these components, their composition typically relies on manual engineering or sample-inefficient reinforcement learning. Existing supervised methods are limited by the lack of MP labels in standard datasets and often fail to generalize across discrete and rhythmic movements. We propose Neural Dynamic Composition (NDC), a supervised framework to compose Dynamic Movement Primitives (DMPs) using only raw trajectories for supervision, given a single example of each primitive. Our approach exploits parallel combinations of DMPs as a differentiable relaxation of categorical primitive selection, and utilizes a warp-invariant loss formulation to identify primitives despite spatial or temporal variations. A staged training curriculum is employed to learn the selection policy and DMP parameters, unifying discrete and rhythmic movements in one framework. Evaluated on 2D handwriting and 3D robotic tasks, our method outperforms baselines on most tasks and metrics. An ablation study demonstrates how each design choice contributes to the system's performance.

Knowledge graph (KG) foundation models (KGFMs) predict links on unseen graphs zero-shot. KGFMs have been pioneered by ULTRA which builds a relation graph from co-occurrences that match motif templates, and passes messages over it to obtain relation representations. More recent KGFMs such as MOTIF and TRIX increase the size of this relation graph, and FLOCK replaces message-passing with probabilistic random-walk ensembles. However, larger relation graphs mean more messages, and more walks mean more sequences to encode, so they improve performance at higher training and inference cost. This cost hinders their use in more complex reasoning tasks, such as complex query answering. To reverse this trend, we design MERIT, the first KGFM with a transformer relation encoder: every relation name attends to every other, and the motifs enter as a soft prior rather than as edges. Over 56 benchmarks, MERIT matches the zero-shot MRR of SoTA KGFMs while being 6x faster than TRIX and over 100x faster than FLOCK at inference. Fine-tuned for complex query answering like UltraQuery, MERITQuery improves over it at matched cost.