Muntaqim Ahmed Raju, Ruizhe Ma, John W. Ramsay, David J. Cornell
原文摘要
Heart rate (HR) forecasting may provide advance information about near-future cardiovascular response, potentially supporting physiological monitoring and workload-management decisions during ruck marching. However, robust model development is constrained by the difficulty of collecting large-scale ruck-march datasets, creating a scarce-data problem in which models must generalize across participants while adapting to individuals from limited personal data. To address this problem, we benchmark population-only training (POT), personalized-from-scratch training (PFST), and population-pretrained fine-tuning (PPFT) for HR forecasting during loaded ruck marching using data from eight active-duty U.S. Army medics across thirteen sessions. LSTM, Attention-LSTM, and CNN-LSTM backbones were evaluated against a persistence baseline at 30 s and 60 s horizons. Across all neural architectures and horizons, PPFT yielded lower mean MAE, RMSE, and endpoint MAE than both POT and PFST. Participant-level paired analysis after averaging across three training seeds showed that, relative to PFST, PPFT reduced trajectory MAE by 2.22-4.30 bpm and endpoint MAE by 1.24-2.88 bpm. All six trajectory-MAE comparisons remained significant after Holm adjustment ($p_{adj}$ = 0.0469), while four of six endpoint-MAE comparisons remained significant after adjustment. At 30 s, persistence achieved the lowest MAE (3.555 $\pm$ 1.274 bpm), while CNN-LSTM PPFT achieved the lowest RMSE (4.669 $\pm$ 1.916 bpm); LSTM PPFT achieved the lowest endpoint MAE (5.093 $\pm$ 2.365 bpm). At 60 s, Attention-LSTM PPFT achieved the lowest MAE (5.573 $\pm$ 1.939 bpm) and RMSE (7.134 $\pm$ 2.260 bpm), while CNN-LSTM PPFT achieved the lowest endpoint MAE (7.553 $\pm$ 2.357 bpm). These results suggest that population pretraining followed by participant-specific fine-tuning may provide a practical strategy for scarce-data wearable HR forecasting.
Yifei Wang, Tsun-yu Yang, Yechen Xu, Hancheng Ye, Qinsi Wang, Baijun Zhou, Tingjun Chen, Hai "Helen" Li 等 10 位作者
原文摘要
Large language model (LLM) agents are rapidly becoming a dominant paradigm for AI applications. Recently, there is a growing demand to deploy large-scale Mixture-of-Experts (MoE) models locally on consumer PCs to balance data privacy, API costs, and inference efficiency. However, unlike datacenter servers with abundant DRAM, consumer devices are tightly constrained in both GPU VRAM and host RAM, forcing massive MoE weights to spill across VRAM, host RAM, and SSD. Under such spilling, we observe that existing systems fail to fully utilize this memory hierarchy, treating the tiers as independent caches and thrashing in inference. We present TIDEWAY, a system for efficient inference of large MoE models on memory-constrained consumer devices. Our core insight is to manage host RAM and GPU VRAM as a unified memory pool. Within this pool, TIDEWAY couples phase-aware memory layouts, which adapt to the distinct access patterns of prefill and decode, with an execution engine that optimizes CUDA Graph execution under expert paging. Our evaluations on representative consumer hardware with frontier MoE models demonstrate that TIDEWAY significantly outperforms state-of-the-art systems. On average, TIDEWAY achieves speedups of 1.97x in end-to-end latency and 2.40x in Time-to-First-Token (TTFT), along with 1.61x higher decode throughput against the fastest baseline. Our source code is publicly available at https://github.com/Afterglow231/Tideway .
Let $\F$ be a finite field. Given $N$ distinct points $\bmα_1,\ldots,\bmα_N \in \F^m$ and associated values $β_1,\ldots,β_N \in \F$, we study the problem of computing an $m$-variate polynomial $f(\vy) \in \F[\vy]$ of individual degree less than $d$, with $d^m \ge N$, that satisfies $f(\bmα_i)=β_i$ for all $i\in[N]$, whenever such an $f$ exists. This is known as the\textit{ multivariate polynomial interpolation} problem. In this work, given $N=d^m$, we present a deterministic algorithm for multivariate polynomial interpolation over an arbitrary finite field $\F$ that runs in near-quadratic time $N^{2+o(1)}\poly(m,d,\log|\F|)$, assuming $d$ is $ω(1)$. This bound holds whenever the associated \emph{multivariate Vandermonde} matrix is nonsingular, has constant co-rank, or admits some special structure. In the general case, we give a randomized Las Vegas algorithm attaining the same near-quadratic running time, provided the finite field is sufficiently large and $d$ grows similar to the previous case. Unlike other multivariate interpolation strategies, we do not assume black-box access to the polynomial. In particular, we are not permitted to query arbitrary input points to obtain their evaluations. Our interpolation result follows by combining Wiedemann's classical approach~\cite{w86} for solving sparse linear systems with known bounds for fast \emph{multivariate multipoint evaluation} over finite fields~\cite{umans08, KU11, BGKM23, BGGKU24}. Specifically, our deterministic interpolation strategy relies on a deterministic variant of Wiedemann's approach, which is more commonly used in the randomized setting.
Reuth Mirsky, Sarath Sreedharan, Nitay Alon, Stefan Sarkadi, Munindar P. Singh, Liz Sonenberg, Peter Stone, Silvia Tulli 等 10 位作者
原文摘要
Agentic AI systems increasingly undertake extended tasks and act on people's behalf, making human-agent interaction central to their effectiveness. Successful interaction requires understanding human perspectives, coordinating complementary capabilities, and maintaining productive relationships within socio-technical environments. This paper examines how foundations developed by the Autonomous Agents and Multi-Agent Systems (AAMAS) community inform these demands. We consider interactions ranging from individual human-agent partnerships to teams and larger groups, where participants differ in their roles, decision authority, capabilities, and objectives. We organize computational approaches around the connected challenges of understanding others and deciding how to interact with them. Theory of Mind, belief-desire-intention architectures, goal and plan recognition, inverse planning, and recursive reasoning provide mechanisms for representing and inferring partially observable mental states. Observer-aware planning, model reconciliation, and decision-theoretic communication connect these inferences to action, explanation, and information exchange. Together with shared autonomy, argumentation, and cooperative inverse reinforcement learning, these approaches offer foundations for agentic systems that account for evolving human knowledge, intentions, and expectations. We argue that progress requires treating human interaction as an ongoing decision problem shaped by uncertainty, interdependence, and potentially conflicting objectives.
We prove the $k$-quasi-planar graph conjecture for topological drawings. For every integer $k\ge 2$, a simple graph on $n\ge 3$ vertices admitting a drawing with no $k$ pairwise crossing edges has at most $1056(k-1)^2(n-2)$ edges. Thus the number of edges is linear in $n$, with an explicit quadratic dependence on $k$. The result allows multiple crossings between the same pair of edges, including edges with a common endpoint. The proof uses the double of a sphere with boundary to represent graph edges by closed geodesics on a hyperbolic surface. The endpoints of geodesic lifts induce partial orders of width at most $k-1$. A stability bound for greedy chain partitions, together with a counting argument over a triangulation, gives the stated edge bound.
Igor Sadalski, Minghao Guo, Benjamin Jones, Wojciech Matusik
原文摘要
Many advances in ML systems came from breaking an assumption that earlier methods took for granted: speculative decoding dropped the assumption that a model needs one forward pass per token. Yet algorithms rarely state their assumptions, so experts and discovery harnesses search over programs, where one edit changes several design choices and gains are difficult to attribute. Searching over the assumptions themselves remains largely unexplored. Writing algorithms as proofs makes their assumptions explicit and turns algorithm design into a search over a small, structured space that an LLM can explore and a proof checker can audit. We formalize 89 algorithms from KV-cache eviction, weight pruning and speculative decoding as proof graphs with explicit assumptions, checked in Lean. Our harness, AutoAssumption, relaxes one assumption of a published method, re-derives and re-proves it, and generates its code. In KV-cache eviction and weight pruning, this search improved on every published algorithm we ran and produced state-of-the-art algorithms. The eviction algorithm has the highest RULER-4K score of the published methods we ran, and the pruning algorithm the lowest perplexity of thirteen on four LLaMA models. At equal budget, the harness scores within a point of OpenEvolve, an open-source implementation of AlphaEvolve, with a lower standard deviation across repeats. Future work will formalize more fields and combine them through the principal assumptions they share.
The standard check for contamination in LLM backtests is simple: compare scores before and after the training cutoff. We show this check is uninformative. Four flagship models fail it on questions they cannot have memorized: every scored question resolved after their cutoffs. The reason is structural. Models legitimately know more about times near their cutoff, so recency mimics leakage, and we prove no passive backtest can separate the two from genuine skill. Measurement, not just detection, requires information from outside the backtest. We supply it in two forms. A known cutoff identifies leakage at the boundary; a matched clean control identifies it globally and yields a leakage-adjusted score. Under a convex-pull model, leakage concentrates on outcomes that surprised the crowd and that training covered well. Partial memorization gives a large part of the full inflation. In twin models with planted leakage, the estimators detect and localize it, track its dose, and return null on clean questions. On frontier models, the estimators detect one cutoff-localized signature. They find no significant leakage at the tested boundaries in five models with positive naive gaps. Backtests need not be discarded; they need one defensible reference.
Agentic retrieval-augmented generation (RAG) interleaves retrieval, reasoning, and answer generation across multiple hops. A retrieval error at hop 1 can surface only as a wrong answer at hop 3, while later retrieval can also repair the trajectory. This paper introduces AgenticRAG-FP, an interventional benchmark for causal failure attribution in agentic RAG. The benchmark injects a certified fault at a specified hop, re-executes the downstream trajectory, and evaluates diagnosers against the known intervention site. Its central question is whether a post-hoc trace still identifies the injected hop after the suffix changes. In the dense Claude Haiku 4.5 structural sweep on three-hop MuSiQue questions, coverage-based diagnosis is 0.91 at hop 1 and 0.00 at hops 2 and 3 (n=43,36,21 failed trajectories), but it predicts hop 1 in 96% of failures, an earliest-hop rule; at hops 2 and 3, an LLM judge and frozen-hop repair localize only 0.25 to 0.48 of failures. In larger content-corruption runs, depth-2 exact-hop accuracy is 0.00 for coverage, 0.49 for an LLM judge, 0.59 for frozen-hop repair, and 0.14 for suffix regeneration (n=51); the judge and frozen-hop repair do not significantly differ (exact McNemar p=0.33).
Yifei Wang, Tsun-yu Yang, Hancheng Ye, Yechen Xu, Cong Guo, Chiyue Wei, Qinsi Wang, Dongting Li 等 12 位作者
原文摘要
Self-hosted LLM agents offer greater control over sensitive data and model customization while avoiding per-request API fees. However, serving concurrent, multi-turn workflows within limited GPU memory makes it difficult to achieve high workflow-level goodput. In particular, as workflows accumulate context, their reusable key-value (KV) caches can easily outgrow GPU capacity, triggering repeated eviction and recovery, known as KV cache thrashing. This process consumes resources and delays workflow completion. Although existing approaches propose various methods, our motivation study shows that the evaluated baselines still incur substantial thrashing overhead or fail to address the workflow-level goodput issue. We present MARS, an agent serving system that coordinates workflow admission and in-engine scheduling to achieve high workflow-level goodput on memory-constrained devices while avoiding KV cache thrashing. Specifically, outside the engine, an admission controller selects root workflows using initial prefix footprint estimates and live KV availability, adapting admissions to memory pressure. Inside, an agent-centric scheduler prioritizes turn completion and coordinates prefill allocation with KV reclamation to reduce disruptive evictions. Across two models and three GPU configurations, MARS achieves 1.59x-7.56x mean end-to-end latency speedups over the fastest baseline. Integrated as the serving backend for OpenCode, Codex, and OpenHands, MARS further improves mean workflow completion time by 1.17x-1.87x. Our source code is publicly available at https://github.com/Afterglow231/MARS_preview .
Modern data centers and cloud computing clusters are increasingly running workloads composed of malleable jobs. A malleable job can be parallelized across any number of cores, yet the job typically exhibits diminishing marginal returns for each additional core on which it runs. Examples of malleable jobs include machine learning tasks, parallelized across multiple GPUs, and large-scale database queries, parallelized across multiple servers. Given the prevalence of malleable jobs, theoretical works have posed the problem of how to allocate a fixed number of cores across a stream of arriving malleable jobs so as to minimize the mean response time across jobs. We refer to this as the Core Allocation to Malleable jobs (CAM) problem. Although quite a few recent papers have been written on CAM, the problem remains challenging to solve optimally. In particular, prior theoretical work has only solved the CAM problem in very constrained cases. This paper takes a new approach to CAM, looking at the problem in the mean field asymptotic regime. In this regime, we are able to optimally solve a much more general instance of CAM. We derive two distinct mean field optimal policies, FW-CAM and GLAM. FW-CAM demonstrates a new intuition: in the mean field regime, individual job sizes are not relevant in finding an optimal policy. GLAM (Gittins-Like Allocation for Malleable jobs) is asymptotically optimal under certain common assumptions, and also serves as a good heuristic outside of the asymptotic regime by adapting dynamically to the current number of jobs in the system. Notably, none of the policies previously proposed in the literature are mean field optimal when jobs follow different speedup functions.
Embodied continual learning requires controllers to infer changing situations, retain prior behavior and reuse motor skills without task labels or known boundaries. The challenge is deciding when to recover earlier behavior and which skills can support new learning. We introduce Situation-guided Primitive Integration for Recurrent Embodiment (SPIRE), a recurrent-spiking system that infers latent contexts from action-outcome evidence and learns context-specific motor actors online. Mature actors become protected sources that combine with a target-context residual to drive recurrent-spiking motor populations. Across coordinate, three-joint and hidden-force control, SPIRE maintained control, autonomously inferred situations, retained prior behavior and accelerated acquisition of an unseen source composition, reducing mean interactions to criterion by approximately 40% relative to matched fresh controllers. Rewinding compatible source actors impaired early acquisition in every matched intervention, demonstrating causal reuse. Against seven recurrent continual-learning and modular-control comparators, SPIRE achieved 2.4 times the mean prior-learning advantage of the strongest eligible task-free alternative selected for each paired stream. These results demonstrate autonomous recovery and causal reuse of motor skills for new learning in continuous embodied control.
We investigate the theoretical aspects of offline reinforcement learning (RL) under general function approximation. While prior works (e.g., Xie et al., 2021) have established the theoretical foundations of learning a good policy from offline data via pessimism, existing algorithms that are computationally tractable (often in an oracle-efficient sense), such as PSPI, only apply to finite and small action spaces. Moreover, these algorithms rely on state-wise mirror descent and require actors to be implicitly induced from the critic functions, failing to accommodate standalone policy parameterization which is ubiquitous in practice. In this work, we address these limitations and extend the theoretical guarantees to parameterized policy classes over large or continuous action spaces. When extending mirror descent to parameterized policies, we identify contextual coupling as the core difficulty, and show how connecting mirror descent to natural policy gradient leads to novel analyses, guarantees, and algorithmic insights, including a surprising unification between offline RL and imitation learning.
Zhouzi Li, Cindy Zhu, Arpan Mukhopadhyay, Mor Harchol-Balter, Benjamin Berg
原文摘要
The past decade has seen a dramatic increase in demand for GPUs to train Machine Learning (ML) models. Because it is prohibitively expensive for most organizations to build and maintain a large GPU cluster, organizations instead choose to rent GPUs from cloud providers. The customer is responsible for devising a policy for (i) deciding how many GPUs to rent at every moment in time to process a stream of ML training jobs and (ii) allocating the rented GPUs among the currently active jobs in the system. Because ML training jobs can be parallelized across different numbers of GPUs, the customer generally has many options for how many GPUs to use for each job. Allocating more GPUs to a single training job will cause the job to complete more quickly. However, the customer pays for each GPU-hour they use, and a training job receives a diminishing marginal benefit from running on additional GPUs. Hence, allocating too many GPUs to a single training job can dramatically increase the overall cost that the customer pays to the cloud provider. This gives rise to a cost-performance tradeoff that customers must balance when running training jobs in the cloud. To balance the cost-performance tradeoff, we develop BOA Constrictor, a new scheduler for ML training jobs which uses a Budget-Optimal Allocation (BOA) policy to squeeze the highest level of performance out of a cloud-deployed GPU cluster given a fixed budget constraint. We explicitly formulate the problem as a budget-constrained scheduling problem and derive the BOA policy which minimizes the average job completion time (JCT) of a stream of arriving jobs subject to the user's budget. For a given budget level, we demonstrate that BOA Constrictor can reduce average JCT by 1.6 times in small-scale implementation experiments and by 2 times in detailed, large-scale simulations compared to state-of-the-art heuristic based schedulers.
Open-weight Vision-Language Models (VLMs) score 85-94% on spatial benchmarks (BLINK Spatial, What's Up CLEVR) yet collapse to chance on simple geometric questions when object, scene, and texture priors are stripped away. We introduce and open-source Sophon, a procedurally generated diagnostic of paired colored point-cloud stimuli across three coordinate systems (1D linear, 2D Cartesian, 2D polar) and four question types (binary, categorical, comparative, quantitative). Across four open architectures (LLaVA-NeXT, LLaVA-OneVision, Qwen2.5-VL, Idefics3), accuracy clusters at 47-52% while frontier models (Opus 4.6, Gemini 2.5 Pro, GPT-5.4) clear 77-83% and humans reach 84-94%. We use linear probes and low-rank adaptation (LoRA) to trace the open-weight failures. Probes of the vision encoder (ViT) and language decoder (LLM) recover Sophon spatial information at 83-85% and 80-86% accuracy, respectively, and the language tower's reasoning circuit answers text-only controls (point clouds as sets of coordinates) at +40 pp over visual stimuli. Language tower per-layer attention attribution localizes visual integration to mid-stack layers. At readout, models elevate the correct answer-set tokens but select among them randomly. We localize the mechanism by tuning individual and grouped components across every architecture's stack, including the vision tower (ViT workspace and reader), the projector, and the language tower (LLM workspace and reader). These experiments demonstrate that this mid-stack workspace holds circuitry capable of organizing the spatial signal so that the frozen reader can decode it. LoRA fine-tuning adjusts that circuitry such that 7B-class VLMs answer Sophon spatial reasoning questions at frontier aggregate accuracy. This work suggests that these models do not have a capability gap for spatial reasoning but suffer from a routing bottleneck that targeted fine-tuning repairs.
Lenstra showed that mixed-integer linear programming can be solved in polynomial time when the number of integer variables is fixed and the number of continuous variables is arbitrary. Mixed-integer quadratic programming (MIQP) can be solved in polynomial time when the total number of variables is fixed. However, since MIQP is NP-hard already without integer variables, the analog of Lenstra's result fails unless P = NP. Hence, polynomial-time algorithms for MIQP with an arbitrary number of continuous variables must exploit additional structure. We study MIQPs with two-stage structure, which underlies Benders decomposition and two-stage stochastic programming. A few linking variables, including all integer variables, are coupled with many blocks of continuous variables that interact only through them, both in the constraints and in the objective. We give a polynomial-time algorithm for MIQP with an arbitrary quadratic objective when the number of linking variables and the block size are fixed, while the number of blocks, and hence the number of variables, is arbitrary. The case with no blocks is MIQP in fixed dimension. None of our assumptions can be dropped. Unless P = NP, the number of linking variables and the block size must both be fixed, the blocks cannot interact or contain integer variables, and large blocks cannot be allowed even if their objective is linear. Unless FPT = W[1], the exponent of the running time must depend on both the number of linking variables and the block size.
RF fingerprinting benchmarks report how well a classifier separates transmitters, not whether that separability reflects the individual unit or its hardware family. We compare a benchmark's real hardware grouping with size-matched random groupings of the same devices, using calibrated, unclipped variational lower bounds under label-blind, device-level cross-fitting on devices never seen in training. On WiSig (OFDM Wi-Fi), the generation grouping exceeded all 999 matched random groupings (permutation p = 0.001, the resolution floor), although no grouping attained a positive certified bound, and the direction survives removing channel equalization. A preregistered replication on a LoRa corpus (chirp spread spectrum, three hardware models), with a front-end frozen on synthetic data, again exceeded all 999 (p = 0.001), and two sensitivity analyses judged by a reversal rule fixed in advance did not reverse the direction. This is a directional replication only: the excess magnitude reflects a miscalibrated null, the chain-rule decomposition leaves a residual of +0.296 bits, and capture order is undocumented. In a same-model corpus of 122 ESP32 radios, a temperature-calibrated held-out bound of +3.338 bits (label entropy 6.931 bits) indicates substantial device-specific identity, not separable from per-device capture-session effects; relying on the calibrated bound was decided after the uncalibrated bound changed sign. On synthetic transmitters a high-capacity convolutional decoder reaches 98.5% of a known 6-bit ceiling, and the audit protocol leaves chain-rule residuals of +0.072 bits (synthetic, fixed physical features) and +0.060 bits (WiSig, audit features). Two further corpora were excluded by preregistered ingest gates. Across WiSig and LoRa, hardware-family groupings consistently produced stronger identity-associated signal than size-matched random groupings.
Digital platforms reward observable claims to build their user base. Because claims are imperfect proxies for people, these programs attract bonus hunters and airdrop farmers, who are usually treated as leakage and screened out. We show that screening can be too aggressive when eligibility requires a separable platform-building input. In a monopolist model, fixed per-claim rewards cannot implement finite farmer activity: rewards below farmer cost induce no entry, while rewards above cost induce unbounded entry. A shared pool can succeed because it makes claims rivalrous. Entry dilutes the pool until the per-claim reward equals claim cost, converting the program into identity-blind procurement. In the baseline model, all farmer-created claims have the same lower cost, while all organic claims have the same higher cost; farmers therefore supply the activity. Organic users receive no rewards but benefit from higher quality, so perfect screening can reduce platform profit and organic-user payoffs. In case when primary organic claims are cheaper but additional organic identities cost more than farmer claims, a small pool attracts only organic claims, whereas a larger pool recruits farmers to expand activity. The mechanism applies to separable inputs (e.g., liquidity, inventory, and data) rather than rewards for regular use or mechanical claim creation. Thus, weak identity changes both leakage and how platforms procure growth.
Karthikeyan B. Kalyanasundaram, Ivan S. Maksymov, Philip H. W. Leong
原文摘要
Ternary large language models (tLLMs) replace general multiplication in dominant linear layers with sign-controlled accumulation, shifting hardware bottlenecks toward packed-weight delivery, reduction, and operator occupancy. PENSA is an end-to-end FPGA accelerator for the BitNet b1.58 2B4T model on an AMD Alveo U50. Two row-partitioned ternary matrix engines consume packed TQ1_0 weights from eight independent high-bandwidth memory (HBM) pseudo-channels, each processing 1,280 ternary weights per cycle through sign selection and integer reduction. The sequencer, attention, and vocabulary-projection units remain resident on the FPGA, while two-stream interleaved scheduling overlaps sequences across the persistent pipeline to form the Penta-Unit. At 104.1~MHz, PENSA achieves 58.09 tok/s at context 128 and 24.60 tok/s at context 2048, corresponding to 2.14-3.14$\times$ the throughput of a same-model Xeon E-2356G baseline. The resulting system demonstrates end-to-end FPGA execution of 2B-scale BitNet topology without modifying its model structure.
Wen-Han Hsieh, Ken Jiankun Zheng, Chia-Chien Chen, David M. Chan, Trevor Darrell
原文摘要
Vision-language-action (VLA) models have made significant progress in general-purpose robotic manipulation. However, sampled action sequences can be ineffective, while evaluating alternatives with computationally expensive verifiers adds latency at each policy query. In this work, we propose HDH, a lightweight framework that uses hindsight from a frozen VLA's own deployments to improve action verification. The framework has three stages. First, we deploy the policy and collect its autonomous rollouts. We then train a compact action verifier to assign each action chunk a predicted task-success score. Finally, the frozen VLA generates action candidates, and the verifier selects the candidate with the highest score for execution. Across four LIBERO-10 tasks, HDH achieves an average success rate of 82.3\% and on three real-world manipulation tasks, HDH improves average success over the baseline policy by 22.6\%.
Cameron R. Jones, Julia Wong, Aaron S. Cohen, James Woolley, Benjamin K. Bergen
原文摘要
People are increasingly using conversational AI systems to make decisions despite concerns around their reliability. How well do people calibrate to AI systems' accuracy in multi-turn conversations? We measure how people's reliance on LLMs' advice changes across repeated interactions, varying whether the AI assistant is helpful (100% accurate), unhelpful (0% accurate), or mixed (50% accurate). Across 4 experiments, we find people learn to trust helpful models more than unhelpful ones (67% vs 21% of the time), but only when they receive external feedback on the models' accuracy. Without feedback, people continue to be persuaded 48% of the time by a consistently unhelpful model (vs 38% for helpful). Rather than gaining trust in helpful models, participant reliance on LLM advisors tended to be high in early trials and only fell with evidence of inaccuracy. Participants trusted LLM advisors more than they trusted human advisors (83% vs 65%, Experiment 4), and LLMs benefited more from having a helpful track record than humans did. In contrast with previous work, we find limited evidence of an advantage for arguing for the truth (versus a falsehood). Truth advantages emerged only when participants were already very confident in their position, and did not emerge at all for static hints. Collectively, the results suggest that people may be swayed by LLM advice, even when it is consistently false, unless they have access to timely ground-truth feedback. Unfortunately, in many domains such feedback may not be available, leading to overreliance on AI.
As particle-based simulations grow in scale, efficient compression of scientific data becomes critical. Neural point cloud compression (PCC) methods excel on computer vision (CV) benchmarks, but do they transfer to scientific particle data? We highlight that scientific particles differ from CV point clouds in fractal dimension: CV point clouds are locally planar, but particles are not. This key difference limits the effectiveness of the octree data structure that underpins most PCC methods. Neural codecs achieve only marginal bitrate reductions over the traditional G-PCC baseline on scientific data, revealing a significant performance gap. Supported by an evaluation spanning domains of cosmology, molecular dynamics, and plasma physics, we outline where the structure of scientific data limits current neural approaches and where learning can help in the future.
Jaeyeon Jang, Diego Klabjan, Veena Mendiratta, Fanfei Meng
原文摘要
Federated learning is an emerging paradigm for decentralized training of machine learning models on distributed clients, without revealing the data to the central server. Most existing works have focused on horizontal or vertical data distributions, where each client possesses different samples with shared features, or each client fully shares only sample indices, respectively. However, the hybrid scheme is much less studied, even though it is much more common in the real world. Therefore, in this paper, we propose a generalized algorithm, FedGraph, that introduces a graph convolutional neural network to capture feature-sharing information while learning features from a subset of clients. We also develop a simple but effective clustering algorithm that aggregates features produced by the deep neural networks of each client while preserving data privacy.
Vision foundation models trained on discretely sampled images achieve strong performance on classification benchmarks, yet whether their representations encode the continuous processes underlying their training data remains unclear. This question is especially pertinent in computational pathology, where we posit that models whose latent representations implicitly capture ordinal disease progression may better reflect underlying biology, support more robust generalization, and enable quantitative analyses of features associated with disease transitions. Using diffusion pseudotime, a method developed to infer developmental trajectories from single-cell transcriptomics, we probe whether foundation models organize disease states along coherent progression directions in representation space. Across four cancer progressions and six models, we find that all pathology-specific models recover trajectory orderings significantly exceeding null baselines, with pathology vision-only models achieving the highest average fidelities. Model rankings by trajectory fidelity on reference diseases predict few-shot classification performance on held-out diseases ($ρ= 0.91$), and exploratory analysis shows cell-type composition varies along inferred trajectories in patterns consistent with known stromal remodeling. Together, these results suggest that vision foundation models can implicitly learn to represent continuous processes from independent static observations, and that trajectory fidelity provides a complementary measure of representation quality beyond downstream performance. While demonstrated in pathology, this framework could be applied to other domains where continuous processes are observed through static snapshots.
Mei Tan, Hyunji Nam, Hari Subramonyam, Dorottya Demszky
原文摘要
Large language models (LLMs) are known to lack diversity in open-ended text generation, a limitation that is especially consequential in education, where homogeneity can narrow the kinds of thinking and disciplinary work students are invited to do. We study diversity in two educational generation tasks: math items and essay feedback. We compare frontier and open-weight models with human-authored reference distributions using lexical, semantic, and domain-grounded measures. For math, we examine the mathematical activities students are asked to perform; for feedback, we examine the revision work requested and the rhetorical strategies used to communicate it. We evaluate direct prompting, two diversity prompting methods, and post-training. Across both tasks, directly prompted model outputs are markedly less diverse than human-authored content--78% higher on lexical (Self-BLEU) and 51% higher on semantic (embedding) similarity--and represent a narrower range of instructional forms. Diversity prompting and post-training increase variation largely through surface features without broadening the instructional repertoire. Models inherit and amplify the most normative routines of human instruction. Overcoming mode collapse in education will require richer instructional data and investment in teachers' creative capacity and professional judgment.
We study the online graph exploration problem, in which an agent must visit every vertex of an initially unknown weighted graph on $n$ vertices and return to its starting vertex. In this model, a vertex's incident edges and their weights are revealed when it is first visited. Despite extensive research, the best known competitive guarantee for general graphs had remained $\mathcal{O}(\log n)$, a bound already established in 1977 for the greedy TSP heuristic Nearest Neighbor. Subsequent approaches, including hierarchical depth-first search (hDFS) and Blocking, achieve better guarantees on restricted graph classes but have logarithmic or worse competitive ratios on general graphs. This raises the question of whether the logarithmic barrier is inherent and Nearest Neighbor is asymptotically optimal. We answer this question in the negative by giving an $\mathcal{O}(\log n/\log \log n)$-competitive algorithm. Our algorithm is a slight modification of hDFS, and we establish its competitive guarantee with a surprisingly short proof.
How well a poetry archive can be searched depends on what the user still remembers, not only on what the model understands. We show this for Classical Arabic poetry, benchmarking nine dense encoders and BM25 over two corpora, three indexing modes and two evaluation protocols. Which retriever wins is governed by how much of a poem's wording the query still carries. On tip-of-the-tongue queries BM25 leads every dense encoder, including a large multilingual embedding model (Qwen3-Embedding-8B, 0.746 vs. 0.717 Recall@5); on abstract queries it falls to eighth out of ten (0.392 vs. 0.748). Editing the wording of the same queries shows that this ranking largely reflects query composition: adding one remembered half-verse raises BM25 by 38 points, rewriting the quoted words by hand leaves BM25 level with the 8B encoder, and BM25's sensitivity to remembered wording is significantly steeper than that of the strongest dense encoder. The same axis controls explanation augmentation, whose effect swings from +37 to -39 points. The mechanism splits by system type: lexical containment best predicts how BM25 responds to explanations, embedding proximity how dense encoders do. Finally, the field's two standard protocols, known-item retrieval and thematic relevance, dissociate on the same systems, a result reproduced by two blind human annotators and by two open-weights judges, one of them under a rubric with no literal-match criterion. A benchmark that reports one protocol, or one query register, can give a materially wrong picture. Code and data are available at https://github.com/Ellzo/what-the-query-remembers.