Science & Mathematics

This paper analyzes graphexon mean field games (GXMFG) on large network limits. By considering edge distributions of a sequence of finite graphs, we define the empirical graphexon measures and study their subsequential weak limits in a graphexon space consisting of all symmetric probability measures on $[0,1]^{2d}$. This concept of network limits extends graphon theory to general, not necessarily dense, networks. To formulate the game using the graphexon limit, we exploit a useful property that a probability measure on a product space is the product of a probability kernel and a marginal distribution. Based on such graph limits and the probability kernel of the associated graphexon measure, we derive the GXMFG equations, and present a detailed analysis for the linear-quadratic case. For linear-quadratic models with finite populations residing over large networks, we use the GXMFG equations to construct decentralized strategies, for which an $ε$-Nash equilibrium property is established.

Safe Multi-Agent Reinforcement Learning (SMARL) aims to learn cooperative control policies under prescribed safety constraints. Al- though existing SMARL methods focus on enforcing safety during policy optimization, their vulnerability remains largely unexplored. This paper investigates inference-time observation attacks against SMARL, where an adversary perturbs the observations of a subset of agents, thereby altering the actions produced by their policies. We propose MinMax Attack, an attack framework that jointly selects adversarial target actions for compromised agents while accounting for the Cooperative Resistance of benign agents. We further derive a bound on the induced safety cost and characterize its relation- ship with the Compromised-Agent Budget and the Perturbation Budget. Experiments across multiple safety-critical multi-agent benchmarks and representative state-of-the-art SMARL algorithms show that MinMax Attack induces substantially higher safety costs than existing attack baselines while preserving task reward, thereby reducing its detectability and achieving both attack effectiveness and stealthiness.

Existing steering benchmarks are built around simplified, isolated generation tasks and evaluate a single behavioral attribute under controlled settings that do not reflect the diversity and complexity of practical deployment. As a result, performance on these benchmarks provides only a limited picture of how steering methods perform in the real world. We therefore introduce RealControlBench, a comprehensive benchmark for evaluating fine-grained steering of LLMs and LLM agents across diverse behavioral targets and interaction settings. RealControlBench extends steering evaluation beyond isolated generations to complex, execution-grounded, multi-turn agent interactions, covering five complementary dimensions: instruction-following control, agentic tool-use decision control, reasoning-performance steering, reasoning-length control, and trajectory-conditioned safety awareness. Rather than measuring steering success alone, the benchmark jointly evaluates target controllability and capability preservation, including instruction following, reasoning performance, and downstream agent utility and safety. It further examines whether steering effects remain specific and robust across prompt variations, domains, and model architectures. Our evaluation shows that existing steering methods remain unreliable across realistic control objectives: they often fail to consistently induce the intended behavior and can substantially degrade the model's underlying task performance. These results suggest that strong steering effects in controlled settings do not necessarily translate into reliable control in more realistic evaluations. By providing a unified, model-agnostic steering evaluation framework, RealControlBench aims to support the development of steering methods that are not only effective but also precise, robust, and useful in realistic AI deployments.

Progress in artificial intelligence is widely assumed to scale with effort: more researchers, papers and investment should accelerate discovery. Theories of technological growth instead locate the difficulty of discovery in the space of ideas. Here we show that neither scale nor space is enough: breakthroughs depend on diverse attention, which collective concentration can deplete. Across 36 years of computer science, we analyze 7.2 million papers, 73,089 peer reviews and 3.7 million researchers and their topic shifts. The field expanded while its attention narrowed sharply after the Transformer. Over half of topic-shifting researchers moved into AI, especially mid-career ones, and reviewer confidence fell by a full point on a 5-point scale. Each successive breakthrough, from AlexNet to large language models, raised the field's carrying capacity for ideas but was followed by steeper exhaustion of those ideas. Bigger breakthroughs capture attention more completely, depleting the diversity their successors may draw upon. Structural biology shows the same cycle after cryo-EM and AlphaFold. Simulations show how well-intentioned and widely used policies such as reciprocal reviewing empower prolific authors and further narrow what the field accepts. We call for institutions that protect diverse attention in artificial intelligence rather than reward concentration on the winning paradigm.

We study the impact of utilizing multimember districts and Single Transferable Vote (STV) as the winner selection method on partisan proportionality. We start with detailed case studies of the upper legislative houses of Illinois, New Hampshire, and New Mexico, building ensembles of redistricting maps with mostly 2-member, mostly 3-member, and mostly 4-member districts for each state. These ensembles are constructed using two methods: building from precincts, and merging currently existing single-member districts. We also compare two ballot generation models for each experiment: a modified Bradley-Terry model and a Plackett-Luce model. We find that our ensembles of multimember districts paired with STV have results which are centered closer to partisan proportionality than when single member districts are used. However, the choice of district size affects partisan outcomes enough that this choice could be exploited by a partisan mapmaker to advantage a particular party. The findings are consistent across all combinations of district construction and ballot generation models within a single state. We then expand our study to 41 different states, obtaining consistently similar proportionality results and allowing us to empirically evaluate the two ballot generation models, as implemented in VoteKit, at a national scale.

Class-Incremental Learning (CIL) aims to learn new classes continuously without forgetting the previously learned ones. Recent works in CIL have significantly advanced across various modalities; however, at the intersection of audio-visual modality, this problem remains underexplored. While representations from multimodal foundational models (MFMs) encapsulate rich static priors, our preliminary analysis shows that these representations struggle in incremental settings. Motivated by this, we propose a framework that generalizes frozen MFMs to audio-visual CIL. Specifically, we leverage their dense audio and visual representations and employ an attention-based contextual guidance module, in which the audio representations serve as task-stable contextual guidance for the visual representations, enabling effective class recognition in the incremental setting. To further mitigate catastrophic forgetting and preserve stability across tasks, we employ feature- and logit-level distillation, which are primarily responsible for the reduced forgetting. Extensive evaluations across various audio-visual CIL benchmarks demonstrate that our approach consistently outperforms state-of-the-art methods. Code is available at https://github.com/avigupta2798/MFM-CIL.

Language-model agents compact their context for the question at hand, but the compacted context must also serve questions that come later. We evaluate compaction on these later questions, using four corpora that annotate the evidence of several questions per text and recorded agent trajectories, and compare every method with random deletion at the same budget. On MultiRC, keeping the sentences that gpt-6-luna (through the OpenAI Decisions API) judges necessary for the current question raises its accuracy on that question by 11.4 points at 25% of the tokens. On the later questions, however, no method that selects sentences for the current question does detectably better than deleting sentences at random. Deleting words instead of sentences does better: LLMLingua-2, which ignores the question, beats random sentence deletion at the same token count by 12.5 points on the later MultiRC questions. It keeps about twice random deletion's share of the evidence's content words, and at equal content-word retention its gain shrinks to 7.3 points (exploratory). On FEVER, label accuracy can overstate what the kept text supports: the gain of sentence selection is no longer detected when answers are scored against the kept evidence, and the 21.0-point gain of LLMLingua-2 is not shown to rest on the kept text. The ranking also depends on the setting: on Evidence Inference sentence selection beats LLMLingua-2 at a nominal 25% budget, and on $τ$-bench trajectories, whose later tool calls need exact identifiers, it beats random deletion while LLMLingua-2 shows no detected gain. Scores on the question a compaction was made for do not reliably show how it serves later ones.

Large language models are deployed globally, yet safety alignment and evaluation remain concentrated in English: harmful requests refused in English can elicit unsafe completions when translated into low-resource languages. We evaluate ten widely deployed LLMs across 79 languages (English plus 78 low-resource) under a translation-bypass threat model, using a metric conditioned on English safety, which separates translation bypass from unsafety already present in English. English safety does not transfer reliably: conditioned on safe English completions, mean bypass rates range from 21.9% to 44.4%. Bypass risk concentrates in a subset of language families, a pattern broadly shared across models. The shared ordering is confounded with translation degeneracy, but its variation across models is significant and survives every control we apply. We also show that per-model bypass rates do not identify a stable ranking: nine of ten models change rank across three conditioning schemes that estimate the same quantity, with no change to any underlying response.

Industrial artificial intelligence (AI) planning systems combine predictive models, optimization logic, controls, and governance rules, but improvement in the complete policy does not show whether one addition made an independent contribution or whether that contribution justifies its burden. We address this attribution problem with an observable evidence standard that separates opportunity to act, executed-plan change, operational consequence, equivalence, uncertainty, and operating scope. Using forecast-to-plan replay on Favorita, M5, and Walmart data, we apply the standard to heterogeneous planning additions. Three limits of single-indicator evaluation emerge. First, a composite policy can conceal negligible component contribution: one control changed only 15 of 681 executed plans. Second, changing execution does not guarantee stable benefit: the evaluated multi-period capability helped in several long-lead settings but was 7.6% more costly under tight capacity. Third, indirect evidence can prioritize review but cannot complete attribution. Independent runs produced different paths in 19 of 599 cases that a rule-based check predicted would be unchanged. Although a lower-cost diagnostic matched the full-evaluation winner in only 55.0% of cases, it selected an option within the $10^{-6}$ cost-equivalence tolerance in 79.5%; the remaining cases prevented the proxy from authorizing acceptance. These experiments establish capability-level contribution evidence; the organization completes the value decision by weighing it against implementation and sustaining burden. The resulting capability evidence record links the observations to validation, deployment scope, fallback, monitoring, redesign, and retirement decisions.

DrugCLIP places small molecules and protein pockets in one space, turning genome-wide drug discovery into large-K retrieval by vector proximity. A campaign fixes over 20K pockets in advance for one library of over 500M molecules. Each pocket keeps K = 500K candidates, 0.1% of the collection, versus 100 in ANN benchmarks. Missed hits never get tested, so mean recall must reach 99% relative to exact search. Large answers make IVF probe every list holding a result and graph search track K or more candidates, so costs grow with K. On our CPUs and GPUs, neither index family reaches the target without losing most of its edge over scanning. SweepANN partitions compressed vectors across GPUs and scans each resident partition. One matrix multiplication scores each block for a whole query batch, reading it once. Several plans perform the K-dependent selection and merging of each query's K answers, each returning an exact top-K under compressed scoring. The window plan retains each query's best K from sampled blocks as its seed, and only rows above each seed's minimum score, shared once across GPUs, enter the buffers. Four Blackwell GPUs (RTX PRO 6000) answer all 26,562 pockets with a resident time of about 61 seconds, excluding library loading, and 99.6% mean recall relative to exact FP32 search. DrugCLIP reports roughly 24 hours with eight A100 GPUs in its own campaign. On 500M rows, the window plan runs 1.41x faster than SweepANN's grouped plan in the same memory. Across 25M-500M rows and K of 10K-500K, all GPU IVF configurations meeting the recall target take 8.2-24x SweepANN's resident time. On DEEP1B's 1B vectors across eight GPUs, against SweepANN's exact full-score scan of uncompressed FP32 rows, the window plan cuts resident time by 5.4x and 3.2x at K = 100K and 500K. It also cuts per-GPU peak memory by 6.0x and 5.3x, losing 1.0 and 0.7 mean-recall points.

For AI assistants to move beyond simple chat interfaces and reliably assist users with applications like Gmail, Venmo, and Slack, we need to improve their tool-use capabilities. However, existing tool use benchmarks are still limited and difficult to scale because they rely on human supervision or are too domain-specific. To make data generation scalable, we propose, TURTLE (Tool Use Related Tasks with Long-Horizon Evolution), an Unsupervised Environment Design (UED) approach to evolve existing tasks using a proposer, a solver, and a feedback generator large language model agents. The proposer is rewarded for difficult yet solvable tasks based on the solver's empirical performance, and the feedback generator summarizes the solver's execution traces for the proposer to exploit. TURTLE produces evolving, self-tuning test suites adaptable to different environment requirements, from supporting user simulators and dense document retrieval to interacting across multiple apps. Empirical results on AppWorld, $τ$-Bench, and AutomationBench using Gemma 4 31B-it show that the evolved tasks target the desired difficulty range (e.g., a 165% increase in the percentage of hard tasks on AppWorld), increase task diversity (in terms of reasoning and simultaneous tool interactions), and elicit a broader range of failure modes than the original tasks.

Optical Character Recognition (OCR) and handwritten character recognition (HCR) in low-resource historical material collections often rely on data augmentation, yet augmentation strategies are typically evaluated almost exclusively through predictive accuracy. This leaves open whether a model's improved performance is accompanied by saliency stability. This question is important in low-resource historical HCR, where recognition is performed on variable and partially degraded letterforms and two models with similar predictive performance may differ in how consistently their saliency patterns respond to such variation. We therefore study saliency-guided augmentation as an explanation-aware strategy for low-resource HCR. Rather than treating saliency maps only as post-hoc interpretations, we use them to define augmentation variants that perturb, preserve, or constrain visually informative regions of character images. We compare classical document-style transformations, random lacunae, and several saliency-constrained lacuna variants. To evaluate the effect of augmentation beyond accuracy, we introduce the Explainability Consistency Index (ECI), a metric that quantifies the stability of saliency patterns under controlled perturbations. We find that saliency-constrained augmentation can improve recognition over classical augmentation, but no single saliency strategy is consistently best. The preferred form of the constraint varies with both architecture and corpus, indicating that saliency is better treated as a mechanism for controlling augmentation than as a fixed augmentation strategy. In some settings, saliency-constrained training improves explanation consistency even when predictive performance is similar. This shows that predictive metrics alone can obscure differences in model sensitivity under perturbation, motivating explanation consistency as a complementary evaluation dimension.

In real-time autonomous field applications, GNSS is often unreliable and inaccurate. A possible solution is localizing against a prebuilt map like a satellite image, as it is georeferenced and already surveyed. In this work, we localize a spinning radar measurement against overhead RGB imagery with a learned observation model. We use a radar representation pretrained by masked reconstruction and align it to features from a satellite image. A particle filter propagated by radar odometry proposes candidate poses, and our observation model provides a likelihood over these poses. We evaluate our method over 31 distinct routes, over 8 datasets spanning both urban and offroad conditions using only ground radar and a coarse local prior. We also examine the robustness over changing weather using repeated traversals over the same route over 13 months, in both urban and offroad conditions.

Multiply-with-carry (MWC) generators are very fast pseudorandom number generators whose quality depends on the choice of a multiplier, and is usually assessed by applying the spectral test to an associated multiplicative congruential generator. We show that the special form of the modulus imposes structural limitations on the lattice structure of every MWC generator, which the generalized MWC generators of Goresky and Klapper mitigate; in our tests, xoring each output with its carry removes them. Since these limitations make the usual worst-case criterion useless, we rank candidates by a weighted geometric mean of their figures of merit, which does not depend on their normalization, and suggest multipliers for generators with 64 to 256 bits of state. Along the way, we correct some errors in the literature. In the end, we propose MWC192X, a 64-bit generator with 192 bits of state that passes all tests we are aware of, and emits an output in half a nanosecond.

Michael O. Rabin was a major founder and leader of theoretical computer science. He died on April 14, 2026, at the age of 94. This paper aims to survey the wide scope and lasting impact of his research.