Transparency has become a central principle in AI governance, yet there is little agreement about what transparency means, for whom, and in what contexts. Drawing on interpretive flexibility within the Social Construction of Technology, this paper examines how different stakeholders understand and experience transparency in generative AI. This study draws from a global survey of 400 end-users, AI builders, and policymakers, complemented by a structured focus group with end-users, grounding the quantitative findings in qualitative insights from the group most underserved by current transparency practice. We present three key findings. First, all three groups share similar understandings of what transparency means and what it should afford, centered on ensuring accuracy and privacy. Second, gaps between the importance of transparency features and their perceived availability are substantially larger for end-users than builders, revealing how power shapes what gets built and who it serves. Only 28% of end-users report that current transparency practices reflect their needs very well, compared to 53% of builders and 49% of policymakers. Third, transparency is situationally constructed, priorities shift with context and task structure, with role-specific frames activating when context demands them and receding when it does not. We argue that meaningful AI transparency requires moving beyond a one-size-fits-all solution, including current emphases on AI disclosure and static documentation, toward pluralistic transparency infrastructures that reflect the needs of diverse users and stakeholders.
We evaluate the extent to which agentic AI models can complete an end-to-end data science project, and how this improves with feedback. A 'writer' model with specific tools creates a complete data science project including downloading, cleaning, and analyzing a dataset, and writing a short paper about it. The project is executed and a 'grader' model evaluates it. The writer can then revise its submission, and this repeats for a total of three rounds. We consider seven models from five developers: Anthropic, OpenAI, Google, DeepSeek, and Z.AI. Each model is both a writer and a grader, so there are 49 writer-grader pairs and we run each pair three times. Of the 147 initial submissions, 107 run and render, and the average score increases from 59 per cent to 76 per cent after the final round of feedback. By writer the extent of the increase ranges from 3 percentage points for Anthropic's Opus 5.5 to 46 percentage points for Anthropic's Haiku 4.5, and by grader it ranges from 13 percentage points for Google's Gemini 3.8 Flash to 32 percentage points for Anthropic's Haiku 4.5. When a model grades its own writing the score tends to be about 9 percentage points higher than the average of what the other models give it. About half of the difference that we see from round to round is within the grader's own noise. Feedback generally results in the highest marks, but even just having the chance to revise, or having the additional budget, in terms of time or tool calls, without any feedback, results in considerable increases for some of the models but not for others. We find that how much such a loop helps depends on which model plays which role, but weaker models generally do not improve stronger models.
Whole-page ranking jointly optimizes advertisements and organic content to balance business outcomes and user engagement under complex layout and business constraints. Existing non-autoregressive (NAR) approaches enable efficient page generation but typically score candidate items independently of previously selected items, limiting their ability to capture order-dependent interactions. Autoregressive (AR) methods address this limitation through sequential conditioning, but incur substantial inference overhead that challenges large-scale deployment. Bridging this quality-efficiency gap remains a challenge in industrial whole-page ranking. We propose Multi-Item Scoring (MIS), a prefix-aware NAR-style framework that incorporates sequential context into efficient multi-item scoring. MIS integrates global candidate interactions, localized prefix conditioning, and position-aware scoring within a shared representation architecture. By computing reusable representations once per request and performing lightweight prefix-conditioned scoring during constrained beam search, MIS captures important sequential dependencies without repeated full-model inference. Experiments on KuaiRand-1K and RL4RS demonstrate that MIS approaches the generation quality of an autoregressive CGA baseline while achieving 4.07x and 3.54x faster CPU decoding, respectively. In a randomized evaluation on a large-scale social-video platform, MIS improves the realization metric by 0.104% and watch time by 0.058%, while reducing ranking-service CPU consumption by 9.73%. These results demonstrate that prefix-aware multi-item scoring can narrow the quality gap between NAR and AR generation while retaining the efficiency required for industrial scalability.
We propose MA-SafeDiffuser, a multi-agent extension of SafeDiffuser that equips diffusion-based trajectory planners with finite-time diffusion invariance guarantees for joint-agent safety specifications. Building on the single-agent construction that embeds control barrier function (CBF) constraints into reverse-diffusion updates, we: (i) formalize joint safe sets as intersections of per-agent and pairwise barrier sets; (ii) derive centralized and decentralized (communication-aware) constrained denoising procedures with provable invariance under mild assumptions; (iii) address local-trap and deadlock phenomena via time-varying specifications and liveness CBFs; and (iv) develop a lightweight benchmarking suite including a multi-agent Maze2D domain. Empirically, MA-SafeDiffuser reduces violation counts relative to unconstrained diffusion baselines, while retaining planning quality. We provide algorithms, proofs, and reference implementations.
AI Agents are becoming digital coworkers that can browse the web, write code, run experiments, and assist humans over extended horizons. Yet existing benchmarks fall short of real knowledge work: many rely on synthetic data or isolated tasks, missing the complex, multi-party dynamics of collaborative projects, where critical information is scattered across conversations, code, and evolving decisions. Even simple workplace questions may require locating relevant channels and direct messages, recovering context from long discussions, inspecting repository state, reconciling stale or conflicting evidence, and synthesizing across sources. We introduce SlackBench, a benchmark for evaluating agents on realistic research and engineering workflows that require joint reasoning over workplace communication and evolving repository state. Grounded in privacy-preserving reenactments of real academic research projects, SlackBench provides reproducible environments with realistic multi-party dialogue, evolving decisions, and corresponding code changes grounded in GitHub repositories and Google Docs. It contains 140 queries spanning heterogeneous-source agentic search, abstention on unanswerable or false-premise questions, goal-directed summarization, and circumstance inference over long conversations. These tasks are not reducible to static corpus question answering: agents must determine which conversations matter, traverse threads, inspect repository state, and synthesize evidence across communication and code artifacts. Across 33 frontier models and agent harnesses, the best system scores around 80% and most agents score below 66%, highlighting substantial headroom for improvement. SlackBench thus provides an extensible framework for evaluating agents in collaborative work environments. The SlackBench dataset is available at https://huggingface.co/datasets/HuanzhiMao/SlackBench .
Large language models (LLMs) are increasingly used to support scientific literature reviews, yet evaluations often examine individual tasks rather than the review process as a whole. We assess six AI tools across four stages: finding relevant articles, interpreting them, synthesizing across them, and proposing future research. The benchmark is a published review of the Spanish plume weather pattern and its 102-article corpus. In the initial searches, five tools collectively identified 57 articles, with the best individual tool finding 42. Institutional library access increased discovery from 42 to 82 articles for Codex and from 15 to 79 for Claude Code. Given the complete corpus of 102 PDFs, three tools correctly identified all 18 articles discussing the Spanish plume substantially, but incorrectly classified 40% to 79% of the 84 passing mentions as substantial. The same three tools produced syntheses evaluated blindly by two domain experts. Although largely accurate, the syntheses provided limited insight and sometimes contained unsupported claims, quotation errors, and exaggerated statements of consensus. Finally, all three tools proposed research directions broader than the Spanish plume literature, although two approached questions identified in the benchmark review. No single tool performed best across all tasks. We provide eight recommendations for researchers using AI tools in literature reviews. Our findings suggest that LLMs can assist with discovering and interpreting scientific literature, particularly when institutional access is available, but cannot replace expert judgment in evaluating evidence or synthesizing research.
Large organizations carry a long tail of infrastructure migrations that are worth doing but rarely urgent enough to compete with feature work, because each one costs significant engineering time per repository. This paper reports on our industrial experience from Automated Uplift, a multi-month initiative applying LLM-assisted migration pipelines across thousands of Bloomberg repositories. We examine two distinct production campaigns: a heterogeneous Python packaging migration (approx. 1,900 candidate repositories) and a uniform C++ build system migration (approx. 6,700 candidate repositories). Together, these campaigns generated 2,333 pull requests, resulting in 729 successful migrations. We detail our hybrid architectural solution which maximizes deterministic processing, confines the LLM to bounded file transformations within a pre-computed context, and incorporates runtime production telemetry (such as execution logs and usage traces). Our findings show that generating migrations turned out to be the easier part; the bottleneck was the cost, for owning teams, of reviewing, testing, and deploying them. These costs were lowered when changes could be verified directly in the pipeline before opening a PR (as with C++ compilation and test validation), and when the population was uniform enough for fixes to carry over from one repository to the next. We conclude by distilling our experience and findings into actionable, transferable patterns for architecting hybrid LLM-deterministic pipelines and scaling code migrations across the enterprise
As agents become more capable, they are moving beyond one-off tasks toward sustained assistance across successive user goals. Such assistance requires agents to discover how unfamiliar environments work, reuse what they learn across tasks, and keep that experience current as environments evolve. Existing benchmarks increasingly study experience reuse and adaptation, but provide limited insight into which parts of accumulated experience agents revise, retain, or lose. To address this gap, we introduce ReScopeBench, a benchmark for the experience lifecycle across 18 interactive environments. It distinguishes structural experience about how an environment is organized and accessed from rule experience about reusable principles and constraints that guide decisions across tasks. To study how experience is acquired, reused, and revised, we propose an experience-first construction approach. It builds task streams around shared experience and introduces silent environmental changes. We evaluate task performance and track changes in saved experience throughout each stream. Across 20 models, we find that retained experience improves task performance, but agents struggle to keep it current as environments change. Even when the required experience is correctly recorded, agents do not always complete tasks successfully. These findings identify selective maintenance and effective use as complementary challenges for long-lived agents.
Large language models understand passengers' driving commands, but as driving agents they reason in generated text for seconds and must infer each action's consequences implicitly from the scene. We present Jev-AD, which pairs Jev, a typed decision model that answers predefined questions with probabilities, not generated text, with a language world model that predicts for every candidate action how nearby vehicles move over the next two seconds. Code turns these predictions into foreseen consequences, which Jev compares instead of inferring them from coordinates. In simulated highway merging (500 car-episodes), this foresight lowers the same decision model's collision rate from 26.0% to 2.4% at 0.25 seconds per decision, and Jev-AD achieves the passenger's intent in 93% of free-form command trials. In 500 replayed real-world NGSIM episodes, foresight lowers the collision rate from 48.2% to 2.8%, and the world model, learned from this traffic, foresees 93% of dangerous actions. Project page: https://ys-qu.github.io/jev-ad-website.
Large language model agents increasingly retrieve tools from large catalogs, yet most retrievers rank each tool independently by request similarity. For multi tool tasks, this recovers the goal matching tool while omitting prerequisite tools required by the workflow. We introduce LIBRA, a framework for Library Induction by Runtime Assembly that turns a flat tool catalog into an executable tool library. An LLM proposes per tool contracts of candidate invocations, reusable outputs and state changes, and success conditions; LIBRA then tests these compositions in a sandbox and retains the relations supported at runtime. Constructed before user queries, the library requires no dependency annotations, tool use trajectories, or task specific training, and supports both one-shot and agentic retrieval. On AppWorld, BFCL, and SciToolEval, LIBRA improves one-shot Recall and NDCG over the same dense retriever by 7.9 and 3.7 percentage points on average. Within ToolReAGt, it yields corresponding gains of 4.5 and 2.9 points after ten retrieval iterations.
For decades, non-player characters (NPCs) have been driven by rule-based systems whose personalities and interactions are scripted before release. Large language models (LLMs) promise NPCs that converse openly, adapt to new situations, and behave in ways developers did not fully author, prompting a wave of research and commercial experimentation. Yet the field lacks a unified picture of what has been built, how well it works, and at what cost. We address this gap with a systematic mapping study of 377 papers from the academic literature, yielding 42 papers analysed against a main research question and ten sub-questions across four thematic axes. We provide the first quantitative characterisation of component adoption, latency, safety coverage, and reproducibility: short-term memory and world-state grounding are near-universal (60% and 69%), whereas long-term memory (31%), reflection (38%), and retrieval-augmented generation (14%) remain rare, and response times span 2.6 to 15 seconds. Finally, we identify six structural barriers to commercial adoption and offer evidence-based recommendations for developers and researchers.
Existing benchmarks for joint text-to-audio-video (T2AV) generation evaluate perceptual quality and audiovisual alignment, but balanced sound-source sam?pling and separate reporting by shot structure remain limited. This makes it dif?ficult to determine whether a specific sound is missing, attributed to the wrong source, or lost across shot transitions. To address this, we introduce CAST?BENCH, a benchmark designed to trace every generated sound back to its in?tended source in multi-shot audio-video generation. The core idea is to formu?late T2AV correctness as an audiovisual contract in which each requested sound must be emitted by the right source, synchronized with the corresponding visual action, and carry the right semantic content. Concretely, the benchmark orga?nizes 900 prompts into 9 sound-source categories, with 100 prompts per cate?gory, spanning speech, animals, instruments, ambient sounds, and other emitter types. The prompts include 439 single-shot and 461 multi-shot narratives, each derived from a structured specification that links every audio event to its emit?ter, subject, and temporal slot. The benchmark provides an 18-metric evaluation suite covering three dimensions: Visual Performance, Content Consistency, and Audio & Semantic Alignment. We further benchmark 12 representative gen?erators and find that visual quality is nearly saturated across leading systems, yet audio semantic scores still differ markedly. Across the nine shared mod?els, multi-shot prompts receive lower aggregate scores under the subset-specific scoring rules and yield different model rankings. Multi-person speech, con?sistent speaker identity across shots, and cross-shot event retention remain the main unresolved challenges. All prompts and evaluation code are available at https://github.com/yucong-liu/CAST-Bench-public.
Genomic-context models have shown that contextualizing protein embeddings can improve predictions of microbial function. We observe that a gene's protein embedding is partially predictable from the embeddings of neighboring genes in the genome, motivating high-ratio masking that requires reconstruction from sparser genomic context. Genomic modeling must also distinguish biological relationships from assembly conventions, since the ordering of separate contigs does not imply adjacency and strand labels depend on the reference orientation. We therefore introduce VenusGenome, a gene-resolution masked autoencoder that combines visible-only encoding with genome-aware attention incorporating contig boundaries, relative strand orientations, and gene positions. We pretrain this model on a curated corpus of over 1.7 million microbial genome assemblies from isolate and metagenomic sources, using frozen protein embeddings as gene-level inputs. Across genome-level phenotype prediction and gene-level tasks, the resulting frozen representations achieve the highest mean performance among the evaluated models on nine of eleven primary metrics. An exploratory gene-input perturbation analysis further links representation sensitivity to experimentally measured fitness defects. Ablations support the contribution of genome-aware attention and show the benefits of masked autoencoding over BERT-style masked modeling, with further gains from higher masking ratios.
Connecting a DNA foundation model to a large language model (LLM) usually requires training a connector, and often the LLM too. We show that a closed-form map is enough for gene-level questions. Our method, the socket, fits one ridge regression (the plug) from frozen Evo 2 features to the host LLM's own representation of the gene's symbol, name, and description at an early layer. The plug's output replaces the hidden state at placeholder positions, and no gradient passes through the host. We test it on GeneHop, a five-way multiple-choice benchmark split by homology. In a single run on the held-out test set, under a frozen analysis plan, real DNA beats shuffled DNA on all four hosts: by +22.2 points [19.8, 25.0] on Qwen3.5-4B, +26.1 [24.3, 27.9] on Qwen3.6-27B, +16.7 [14.2, 19.7] on Mistral-7B, and +18.4 [16.2, 21.0] on OLMo-2-7B. All four survive Holm correction, every null control stays within two points of the sham level, and the development results agree within 2.3 points. A layer rule that uses no task items was selected on three hosts; it predicted the working layer of the fourth before any DNA-plug scoring on it. A text-only linear map lets one plug feed any other host, at a cost of at most 0.9 points on test. The socket conveys gene identity: nearest-neighbour retrieval recovers 72-84% of the effect, but the plug beats it on every host by 3.0-5.2 points and keeps most of its effect without training homologs.
Tiered inference distributes large language model computation across a hierarchy of devices, from on-device and edge hardware to the datacenter, and is increasingly offered as the route to cheaper, lower-latency, and more private serving. Most such systems orchestrate frozen model checkpoints, and most decide what to resolve locally and what to escalate using a confidence signal produced by a forward-deployed model. We give a closed-form condition for when no threshold on that signal makes a tiered cascade worthwhile relative to sending every query to the backstop. Modeling the routing decision as a binary test and separating local from escalation cost through an architecture cost ratio $\kappa$, we derive a feasibility condition bracketed by a signal-free floor, below which a cascade is feasible at any signal quality, and a perfect-signal ceiling, above which none is. Read along the $\kappa$ axis the condition inverts: where local compute is billed, cost binds, matching a recent empirical finding; where it is sunk, on decentralized hardware, feasibility becomes unconditional and signal quality instead sets what the cascade achieves. We turn the condition into three design requirements that existing trained-in primitives satisfy only in isolation, and we propose, as future work, a co-designed architecture and a protocol to test it. The condition and the inversion hold independently of that architecture; they are the paper's contribution.
Large language models perform strongly on standard medical QA but degrade on harder reasoning tasks, where multi-agent debate often fails to beat strong single-model prompting. We argue that a key limitation is that existing systems select experts and retrieve knowledge from the question alone, rather than from gaps that emerge during reasoning. We propose challenge-aware adaptive recruitment: an external CHALLENGER inspects the evolving reasoning trace and identifies specific factual or expertise gaps. These signals trigger targeted knowledge retrieval or mid-discussion specialist recruitment. On the MEDAGENTSBENCH hard subset (862 questions across nine biomedical datasets), upfront retrieval provides little benefit, while challenge-aware in-between retrieval performs better; on a multi-specialist backbone it improves accuracy on 8 of 9 datasets and adds 6 percentage points over Challenger Only. Ablations indicate benefits from both retrieval timing and gap-specific query signals, and a live web-search experiment shows that the timing advantage persists across retrieval backends. For specialist recruitment, self-triggered adaptation is unreliable, whereas Challenger-triggered recruitment raises GPT-4.1-mini MedQA accuracy from 46% to 54%, with positive direction on all tested generalization datasets. Naively combining retrieval and specialist recruitment degrades performance, highlighting the need to coordinate adaptive intervention channels. These results show that when and why external resources are recruited is a key design choice in multi-agent medical reasoning.
Agentic artificial intelligence (AI) systems powered by large language models (LLMs) execute as autonomous multi-step pipelines that dynamically invoke tools, integrate observations into memory, and iterate until a task is resolved or a resource budget is exhausted. Unlike classical workflows with fixed graphs, the aggregate compute, bandwidth, and end-to-end latency demands of such pipelines are unknown at dispatch time: they emerge at runtime from intermediate observations and stochastic token sampling, yielding heavy-tailed, multi-modal random variables. This unpredictability forces infrastructure orchestrators into a dilemma: under-provisioning risks mid-execution throttling and deadline violations, while over-provisioning wastes scarce edge or cloud capacity. Existing schedulers require a fully known pipeline graph, while edge resource-management methods rely on parametric or stationary demand models, and neither fits the unknown, non-stationary profiles of agentic pipelines. We formalize this setting as the uncertainty-aware agentic orchestration problem, where the goal is to reserve, at dispatch time, compute and bandwidth allocations that cover actual pipeline demands, and to verify that the pipeline can meet its deadline, each with a prescribed probability. We then propose Conformal Reservation for Orchestration (CRO), an online algorithm that translates a sliding window of past execution traces into dispatch-time resource allocations with finite-sample, distribution-free coverage guarantees via conformal prediction. Simulations over heterogeneous task mixes with multi-modal and heavy-tailed profiles show that CRO is the only evaluated method that simultaneously satisfies target probabilistic coverage across all risk levels while, among coverage-feasible methods, attaining the lowest over-reservation, roughly halving the resource waste of the Chebyshev-bound baseline.
Tool descriptions tell an agent what an action can do, but they do not necessarily reveal whether that action is healthy or suitable at the moment of use. Execution harnesses often possess this information through monitors, traces, and recent failures while the model choosing the next action does not. We call this gap operational blindness. We introduce Agent Harness Awareness, a trust-separated design in which the host supplies a small, fresh, verified state snapshot before a consequential action while retaining measurement, enforcement, evaluation, and final dispatch authority. We test its tools-and-network instance with ToolRoute, a controlled isolation benchmark that holds the task and candidate semantics constant and varies the operational information available to the model. In a 216-decision ablation across GPT-5.6 Sol, Claude Sonnet 5, and Gemini 3.7 Flash, verified tool-health state reduced mean observable policy regret by 96.9% against a same-shape neutral control (4,252 ms to 131 ms) and raised single-dispatch completion from 69.4% to 95.8%. In a separate 27-decision end-to-end replication with live local TCP/HTTP execution, OpenTelemetry client spans, and Toxiproxy-injected faults, verified state reduced mean regret from 8,956 ms to 0 ms (zero excess observable operational cost) and completed the live action in 9/9 model-by-fault cases versus 5/9 for the neutral control. These studies establish the tools-and-network instance of Agent Harness Awareness: selected host telemetry materially improves pre-dispatch operational choice. Runtime state, hardware pressure, economics, retry-aware recovery, and heterogeneous selection remain open empirical extensions of the same architecture.
Multi-agent language-model systems often add a reviewer to improve an upstream agent's decision. Yet a reviewer can both repair errors and replace decisions that were already correct. We study this as a problem of institutional jurisdiction: conditional on the same proposal, task information, model, and review output, when is the reviewer authorized to make its recommendation binding? We compare broad override authority with an evidence gate that permits replacement only after the reviewer produces a mechanically valid record of a binding-constraint violation and a compliant repair. Across two locally run open-weight models, 96 spatial-policy tasks, and a 96-task portfolio-transfer benchmark, the gate sharply reduces corruption of correct proposals while retaining most repairs of violations. The directions of this tradeoff are partly implied by the rule; its magnitude depends on reviewer transitions and upstream proposal states. A prospective validation then gives the reviewer an explicit choice to retain, replace, or escalate. This eliminates correct-proposal corruption for Qwen, but Mistral still replaces all 48 correct proposals; Qwen also declines every welfare-only repair that Mistral frequently attempts. Thus interface design changes reviewer behavior but is not a model-independent safeguard. The result is not that restricted review is universally better. It is that reviewer competence, action interface, and binding jurisdiction jointly determine both the correction domain and the corruption surface.
Robot foundation models have shown remarkable performance in robot control, yet adapting them to a new task still relies heavily on teleoperated demonstrations. Egocentric human videos offer a scalable substitute, but overlaying rendered robots on human videos or retargeting human poses into robot actions leaves significant visual or action gaps for high-DoF robot embodiments. To address this limitation, we propose RoboTryOn, which treats human-to-robot translation as virtual try-on: try-on keeps the person and swaps the garment, RoboTryOn keeps the demonstrated motion and swaps the human for the robot. Specifically, RoboTryOn converts a human video into a video of the target robot specified by a reference image, while preserving the original task, objects, and background. To maintain visual consistency, we translate selected keyframes autoregressively and use them to guide video synthesis with motion conditioned on the human hand skeleton. Additionally, we use an inverse dynamics model (IDM) that incorporates depth features to infer action labels from synthesized videos for policy learning. Our experiments across multiple humanoid embodiments show that policies trained with our synthetic data improve the average success rate on novel tasks, including challenging bimanual manipulation, from 2% to 60%, without any real robot demonstration of those tasks. Moreover, training with our synthetic data matches the performance of policies trained solely on teleoperated demonstrations at approximately a third of the collection cost, demonstrating the cost efficiency of our approach.