让下一篇重要论文，不再错过。

As LLMs improve, data agents are emerging as a new interface for analytics, while semantic query processing engines use LLMs to operate over structured and unstructured data. Existing benchmarks largely evaluate agents and semantic query engines in isolation, leaving it unclear whether agents can decide when and how to use semantic operators in end-to-end analysis. We introduce AvalancheBench, a multimodal enterprise analytics benchmark that generates observed data from latent world specifications and evaluates data agents on their ability to recover that world. The benchmark's 3 validated tasks contain 58 questions spanning text, images, and documents. Questions are provided together as a multi-part analytical brief and may share intermediate results, exposing the avalanche effect: errors can propagate through legitimate dependencies, and inappropriate reuse can cause otherwise unrelated questions to fail. AvalancheBench measures analytical fidelity, monetary cost, and capability-adjusted token cost (CATC), a new metric that weights token use by model capability to separate technical efficiency from provider pricing. We evaluate 5 model-harness configurations across 3 agent harnesses using Snowflake Cortex AISQL as the engine, with the best configuration reaching macro analytical fidelity of 0.82. Agents fail primarily in constructing the analysis: failing to recover facts that require semantic computation costs 45.1% of lost points, yet incorrect operator outputs account for only 2.5%, versus 36.3% from agents misusing or avoiding semantic operators. Misinterpreting requirements costs the most, at 51.5%. Errors also spread between questions: 29% of lost points come from reusing another question's results. Finally, monetary cost and CATC produce different Pareto frontiers, distinguishing economically attractive configurations from those that are technically resource efficient.

The dilation of a geometric graph measures how much longer the path between pairs of points becomes when restricted to graph edges, rather than following the direct path through the ambient space. The minimum dilation tree of a point set is the spanning tree with minimum dilation, where edge lengths in the tree are given by distances in the ambient space. In the Euclidean plane, computing the minimum dilation tree is NP-hard, but no hardness of approximation result is known. On the other hand, the minimum spanning tree is an $(n-1)$-approximation to the minimum dilation tree, but no asymptotically-better approximation algorithm is known for general point sets in the Euclidean plane. We give the first sublinear approximation algorithm for the minimum dilation tree in the Euclidean plane. Our approximation ratio is $\tilde{O}(n^{14/15})$ and our algorithm runs in polynomial time. This resolves an open problem proposed by Eppstein in 1996.

Automated identification of individual animals is moving from small curated studies toward shared benchmarks and reusable foundation models. This progress is useful, but current evaluation leaves the evidence behind a match unresolved, i.e., high retrieval accuracy does not show whether a model used animal features or identity-correlated shortcuts. We address this with Counterfactual Foreground Retrieval (CFR), an evidence-aware evaluation protocol. CFR evaluates the same query-gallery task under different views: full image, foreground-only, and shortcut-only. The full-image measures standard retrieval performance; the foreground-only measures foreground-supported retrieval; and shortcut-only view measure how much identity signal is recoverable from background, or silhouette. This turns a retrieval score into an evidence profile: how well the model performs on the original images, how much of that performance remains when only the animal is visible, and how much identity signal remains recoverable when the animal is removed. We release around 0.5M masks for 34 datasets and code for mask generation and CFR evaluation.

Code-as-policy agents often use execution feedback to retry failed programs. Additional sampling and task guidance can also improve success, leaving the contribution of failure-specific feedback unclear. We study this question in simulated robot manipulation. A fixed lookup selects feedback templates using failure labels and the API variant. A frozen code model generates a fresh program after each scene reset. On the same 252 initial failures, Qwen2.5-Coder-7B recovers 85.3% within five retries, compared with 68.3% for resampling with the same retry cap. The advantage persists under fixed total-token budgets. Feedback from other episodes and fixed task guidance yield similar recovery point estimates, with no detected benefit from matching the coarse failure label. In a separate study, a fixed guided prompt used from the first attempt achieves an observed overall success rate similar to that of template-based feedback. Template-based feedback also lowers recovery on a nut-insertion task requiring a rim grasp, consistent with an incompatible center-grasp instruction. Evaluations of robot self-repair should include resampling, mismatched feedback, and concise, task-compatible static guidance as baselines.

We investigate whether a multiscale tensor-network architecture provides a useful inductive bias for reconstruction-based anomaly detection in collider jets. Jets are produced by a branching cascade, so their internal structure is naturally organised across angular and momentum scales. This motivates an autoencoder that compresses information hierarchically and can reorganise short-range correlations before coarse-graining. Guided by this picture, we formulate a Multiscale Entanglement Renormalization Ansatz (MERA)-inspired autoencoder acting on ordered jet constituents. To our knowledge, a MERA-inspired autoencoder has not previously been proposed, and this architecture has not been explored in collider anomaly detection. We compare this architecture to a dense autoencoder, the corresponding tree-tensor-network limit, and classical baselines within a background-only reconstruction framework. The paper is organised around two main questions: whether locality-aware hierarchical compression is genuinely supported by the data, and whether the disentangling layers of MERA contribute beyond a simpler tree hierarchy. To address these questions, we combine benchmark comparisons with a training-free local-compressibility diagnostic and a direct identity-disentangler ablation. In the nominal benchmark, the MERA-inspired model attains a higher mean AUC than the dense autoencoder at all three latent dimensions while using roughly one-third of its trainable parameters. The locality-preserving multiscale structure is well matched to jet data, and, at the strongest compression bottleneck, the trained disentanglers move substantially away from the identity and are accompanied by a positive mean performance shift. Overall, the study supports locality-aware hierarchical compression as a useful and parameter-efficient inductive bias for reconstruction-based jet anomaly detection in this benchmark setting.

Safe reinforcement learning (RL) aims to learn policies that optimize rewards while satisfying constraints. Predominant approaches rely on soft-constrained policy optimization, which has achieved empirical success but does not provide formal safety guarantees for the learned policy. In contrast, methods with strict guarantees typically rely on explicit certificate functions, whose construction requires the direct synthesis and verification of control-invariant sets, a process that scales poorly with state dimension and often yields overly conservative behavior. In this paper, we present the Provably Safe, yet Scalable RL (PS2-RL) framework, a novel two-phase architecture for learning provably safe policies in a scalable manner, designed to overcome the key bottlenecks of prior methods. Rather than explicitly computing invariant sets, PS2-RL leverages a learned backup policy to forward-integrate the system dynamics, generating an implicit control-invariant set online. In the first phase, the backup policy is trained with the safe-arrival value function to achieve rapid, safe recovery to the certified base set for invariant-set construction. In the second phase, an RL policy is trained end-to-end through a differentiable projection layer that strictly enforces the safety guarantees induced by the learned backup policy. By maximizing the volume of the implicit control-invariant set in the first phase, the resulting PS2 policy from the second phase is performant and scalable, while maintaining provable safety. Crucially, PS2-RL imposes no restrictions on the underlying RL algorithm and can be plugged into any existing training pipeline. We establish theoretical guarantees for the proposed framework and evaluate it on robotic control tasks with state dimensions up to 10, a regime in which prior provably safe RL methods struggle or become impractical.

Hierarchical text embedding models offer variable-length representations, while quantization reduces numerical precision; both aim to improve computational efficiency. This work is an empirical benchmark study of how these two orthogonal, already-established techniques interact. We evaluate 38 tasks from the Massive Text Embedding Benchmark (MTEB) using four transformer-based embedding models, covering 1D and 2D hierarchical representations combined with 32-bit, 8-bit, and binary quantization, for 168 configurations per task. We report each efficiency dimension separately: theoretical per-vector storage and end-to-end evaluation time. Under a Friedman test with Nemenyi post-hoc analysis, int8 quantization is statistically indistinguishable from float32 for pair classification and re-ranking and costs at most 1.5 percentage points overall, whereas binary quantization is significantly worse across every task category. Halving the embedding dimension once is not statistically detectable above $m=16$; two successive halvings always are. Combining an 8x dimensionality reduction with int8 yields a 32x reduction in per-vector storage at 94.4% of baseline effectiveness. Because evaluation times were measured once, we treat them as indicative: we find no evidence that quantization speeds up evaluation, whereas sampling from intermediate layers shortens it substantially. Retrieval degrades far faster than every other category, so we report findings per task category.

Recursive self-improvement requires AI systems to improve the models and infrastructure used by subsequent systems. LLM compression offers a concrete efficiency objective: retain a pretrained model's behavior within a smaller model storage budget. We study whether coding agents can carry out this step autonomously. We formulate compression as a terminal task in which an agent receives a source checkpoint, storage limit, inference environment, and time budget, then constructs compressed weights and executable inference code. The agent controls its algorithm and experiments; perplexity feedback guides search, and independent verification checks the final model. Across six agents and six model--size settings, 29 of 36 submissions pass verification. Selecting the lowest-perplexity agent design per task yields an average 3.4% perplexity increase over BF16 and 77.33% downstream accuracy, compared with 79.56% for BF16. Recorded experiments reveal how agents achieve these outcomes and where improvement breaks down. Agents lower PPL in 21 of 24 revisions at nearly fixed storage; complete-model feedback reverses 17 of 79 prior local rankings; and generation and inference tests expose failures that likelihood alone misses. The results identify representation design, experimental judgment, and capability validation as central to agent-driven model compression, a measurable component of efficiency-oriented self-improvement.

A fault-tolerant (FT) connectivity oracle for an $n$-vertex graph $G=(V,E)$ determines whether $G- F$ is connected for any query set $F\subseteq E$. When the number of failures is unrestricted, prior constructions exhibited a sharp contrast between randomization and determinism: randomized oracles use only $\widetilde{O}(n)$ space [Duan and Pettie, SICOMP 2020], whereas no deterministic construction was known to improve upon storing $G$ explicitly, requiring $Θ(n^2)$ bits in the worst case. The promise that $F\subseteq E$ is essential for subquadratic space: allowing arbitrary queries $F\subseteq V\times V$ forces quadratic space. For queries containing at most $f$ failures, Khanna, Konrad, and Putterman [SODA 2027] recently constructed deterministic oracles using $\widetilde{O}(\sqrt{f}n)$ space that also report approximate distances with polylogarithmic stretch. Their result improves upon the classical $O(fn)$ space bound for deterministic connectivity oracles [Pătraşcu and Thorup, FOCS 2007], but offers no space savings over storing $G$ when $f=Θ(n^2)$. We present the first deterministic connectivity oracles with subquadratic space for failure sets of unrestricted size, using $\widetilde{O}(n^{7/4})$ space. We also establish the first polynomial separation between deterministic and randomized connectivity oracles, proving that every deterministic oracle requires $\widetildeΩ(n^{3/2})$ bits in the worst case. For queries containing at most $f$ failures, we obtain deterministic connectivity oracles using $\widetilde{O}(f^{2/5}n)$ space, improving the dependence on $f$ in the bound of Khanna et al., and prove a lower bound of $\widetildeΩ(f^{1/4}n)$ bits.

We introduce LongBanana, to our knowledge the first expert-designed and expert-verified benchmark for long-context multi-reference image synthesis in the age of agent-driven image generation. The task requires retrieving dispersed visual evidence, rejecting distractors, binding entities and relations, and composing one coherent final image from multiple references. LongBanana contains 90 carefully designed tasks organized by a 3-by-3 taxonomy; each task includes about 32 reference images, 20 constraints, and 50 evaluation items, yielding 4,670 expert-specified atomic evaluation items in total. Its graph-grounded evaluation design instantiates each task with reference images and an entity-relation graph, then converts node, edge, and whole-graph obligations into checklist items for Content Fidelity, Relational Binding, and Global Constraint Satisfaction. Reviewed expert labels include a 15% double-labeled subset with 89.4% exact agreement and 0.83 quadratic weighted kappa. We further introduce LongBanana-Gen, an agent-harness framework for long-horizon image generation; experiments show that minimal agent harnesses improve over direct generation, while a ReAct-Plan-Exec-Reflect (RPR) workflow yields harness-dependent gains by strengthening evidence grounding and repair. Together, LongBanana provides an auditable testbed for measuring dense visual evidence use and a shared target for developing stronger agent-driven image-generation systems. Code and data are available at the project page: https://long-banana.com/.

Computer-use agents (CUAs) have recently achieved superhuman performance on several long-horizon benchmarks. However, their computational cost limits their use in long-horizon tasks and real-time domains such as video games. A key bottleneck is representing each screenshot as thousands of visual tokens, which accumulate into long contexts as agents reason over extended interaction histories. We observe that most of these visual tokens are either semantically irrelevant to the agent's next action or temporally redundant across screenshots. Building on this observation, we introduce Visual Token Routing (VTR), a joint training framework for learning sparse visual representations that preserve the information needed for action prediction. In VTR, we train the vision-language model together with a lightweight spatio-temporal router that examines each image region in the current and previous screenshots and decides whether its visual token enters the language model. We optimize the router using an adaptive loss to keep the number of visual tokens as small as possible while minimizing performance degradation. With just 2,000 training steps, VTR matches the performance of the full-resolution model on OSWorld, CUA-World, and video game action prediction, while using 2.5-5.8x fewer visual tokens. These savings reduce inference cost by up to 3.3x and speed up prefill by up to 2.9x. VTR also outperforms lowering the image resolution and other baselines for dropping tokens, by up to 12.8 points on OSWorld. Overall, VTR enables computer-use agents to process only the visual tokens they need, allowing them to be used in long-horizon tasks and real-time domains at a fraction of the computational cost. Code and models are available at https://pranjal2041.github.io/vtr

Video generation models synthesize realistic motion, but can they respond correctly to changes in initial geometry? We introduce GeoFlip, a video generation benchmark centered on physical outcome flips induced by geometry. With shared physical settings and textual prompts held fixed, GeoFlip applies controlled interventions to support extent, object placement, and mechanism configuration, constructing paired scenes with contrasting reference events, such as remaining stable versus falling or tipping. These pairs test whether models translate geometric differences in the first frame into correct physical evolution. The benchmark comprises 482 scene pairs and 964 reference videos across six physical categories, with reference outcomes established through procedural scene construction, physics simulation, and numerical validation and verification of physical events. We evaluate 16 models for image to video generation along three dimensions: event correctness, structural integrity, and similarity to reference videos. We examine evaluation consistency with independent vision language judges. On the 442 conditions shared by all models, the highest event correctness averaged across tasks is 27.98%, revealing substantial limitations in physical generation conditioned on geometry. Model strengths vary across physical mechanisms, and the three evaluation dimensions capture distinct capabilities. Typical failures include ignoring geometric changes, generating incorrect object motion, bending rigid objects, and introducing additional components. Rank correlations between judges reach 0.906 for event correctness and 0.929 for structural integrity. By combining controlled geometric interventions, verifiable physical references, and multidimensional evaluation, GeoFlip assesses progress from visual realism toward physical consistency conditioned on geometry.

Common methods for monitoring features in large language models (LLMs) such as linear probes and sparse autoencoders (SAEs) assume that features maintain a consistent representation across contexts. LLMs, however, can flexibly adopt different perspectives or "personas" that suggest different underlying systems of conceptual relations. In this paper, we examine whether prompting an LLM to adopt a persona alters the representational geometry of the model's latent space. We extract representations of over 9,000 common words and 33 semantic axes from models adopting liberal and conservative political personas and compare these against two neutral baselines. We find that persona conditioning systematically alters both the projections of words onto semantic axes and the alignments among the axes themselves in directions consistent with the specified ideology. We then test the implications of this reorganization for two safety-relevant features: truthfulness and morality. Linear probes, SAE-based classifiers, and classifiers built from manually selected word representations, all trained without a persona, produce divergent results when applied to models conditioned on political personas, producing politically slanted ratings of truth and morality on controversial statements. These findings suggest that internal representations, as they are conventionally extracted, can exhibit different geometric relations when the model adopts personas, potentially undermining monitoring efforts.

Faithfulness requires that algorithmic explanations accurately reflect the model's behavior. Yet existing operationalizations of faithfulness can disagree about whether the same explanation is faithful, leaving developers and auditors to choose among conflicting evaluations. In this paper, we investigate disagreement among evaluation criteria and offer a practical path forward for evaluating explanation faithfulness. We first identify how operationalizations of faithfulness differ. We introduce a faithfulness functional that locates their disagreements in what an explanation claims about the model, which comparisons bear on that claim, and how agreement across those comparisons should be aggregated, and we reconstruct existing evaluation criteria within it. We argue these reconstructions support understanding faithfulness as a common though incompletely specified relation between the model and explanation, so the differences reflect choices left open by the general requirement. Justifying an operationalization then comes down to justifying these choices, which, on our view, draws on explanatory, institutional, and normative considerations. We support this argument with a case study of adverse action notices, reconstructing the evaluation implicit in an industry-proposed procedure and identifying which of its choices the setting justifies and which remain open. Our work gives auditors and practitioners a way to navigate disagreeing evaluations of faithfulness. It also explains why the literature should expect multiple operationalizations, while showing how concrete settings can narrow the choice among them. More broadly, our analysis shows how a shared requirement of faithfulness can support contextually justified evaluations despite disagreement among existing metrics.

A central yet under-explored question in Multimodal Large Language Models (MLLMs) is where to connect---how visual information from a hierarchical vision encoder should be wired into the layer-wise semantics of a large language model. Mainstream MLLMs inject a fixed single-layer visual representation into the LLM and treat all LLM layers as a single consumer, ignoring that different layers demand different visual granularities. Recent multi-layer fusion methods either compress hierarchical features on the encoder side or rely on predefined sparse or hierarchical layer-to-layer connections, and therefore stop short of modeling adaptive cross-layer interactions between the ViT and LLM hierarchies. We address this gap with DGP (Dynamic Gated Pathways), a routing mechanism that establishes all-to-all connections between every ViT layer and every LLM layer, with each LLM layer dynamically gating its preference over visual layers conditioned on its own semantic state. Built on top of LLaVA-1.5-7B and 13B, DGP delivers consistent gains over the LLaVA-1.5-7B baseline (e.g., +4.2 on SQA$^{I}$, +3.3 on MM-Vet, +3.1 on VizWiz and LLaVA$^{W}$) and surpasses prior multi-layer visual fusion methods on most benchmarks. Layer-wise routing analyses further show that different LLM layers do select different visual granularities on demand, supporting both the effectiveness and interpretability of the proposed dynamic gated pathways. The code is available: https://github.com/YingyingYan1/DGP.

Grain growth during heat treatment strongly affects the microstructure of polycrystalline materials and, consequently, their mechanical properties. Although numerical methods such as front-tracking or front-capturing simulations can accurately describe grain growth, their high computational cost limits their use for long or non-isothermal thermal cycles. In previous work, grain growth was successfully predicted under isothermal conditions by treating grain size distributions as time series and forecasting their evolution using recurrent neural networks, allowing complete grain growth trajectories to be generated in seconds rather than through costly PDE-based simulations. In the present work, this forecasting framework is reformulated to handle non-isothermal conditions by explicitly incorporating thermal information into the model architecture. Grain size distributions are encoded using a convolutional neural network, thermal descriptors are embedded through a dedicated multilayer perceptron, and temporal evolution is predicted using a recurrent network. Instantaneous temperature and thermal rate are provided at each time step, allowing the model to adapt its predictions to changing thermal conditions. The framework is trained and evaluated on datasets generated from validated front-tracking simulations across a wide range of non-isothermal thermal paths. The results show that the thermally conditioned model accurately reproduces grain growth dynamics under complex thermal histories, achieving, in k-fold cross-validation over all simulations, mean relative errors of 2.99% (one-hour cycles) and 4.51% (four-hour cycles) with coefficients of determination exceeding 0.99, and mean relative errors between 1.91% and 14.66% on ten simulations following four unseen thermal histories of increasing complexity.

Large language model (LLM) coding agents develop software through experimentation, offering an opportunity to design the serving systems supporting their own operation. Effective serving requires coordinating implementation and operating decisions under session dependencies, latency constraints, and correctness requirements. We introduce ServerBench, a benchmark for agentic design of complete, specialized LLM serving systems. Agents experiment against a common objective, dollars per dollar: latency-qualified token value per dollar of GPU cost. Stateful replay preserves session dependencies, and independent verification checks the submitted system. Across 11 coding-agent models and three deployments, Claude Sonnet 5.5 achieves 2.97 dollars per dollar on Gemma E4B, $4.87\times$ the vLLM reference; 22 of 33 runs yield verified systems, and 13 exceed the strongest baseline. Our analysis identifies three findings: high-value designs coordinate capacity improvements with load selection; agents interpret performance feedback through workload-relevant experiments; and reliable delivery requires verification of the execution paths introduced by optimization. Overall, ServerBench supports systematic study of how agents translate service objectives and feedback into effective, verified serving systems for given task specifications.

Ask SD3.5-Medium for twelve of an object and then, from the same starting noise, for twenty-five. The picture keeps its layout, but the count changes about as much as it would under a new random seed, and in no consistent direction. We study what number words do once a request passes about ten objects, in five open text-to-image models. Holding the object and the starting noise fixed, we change one thing at a time, and people who never see the request count 3,088 of the images. Every model follows the number for small crowds, but only up to a point, its reach: SD3.5-M and FLUX.1-dev stop following it soon after ten, while FLUX.2 [klein] and Qwen-Image follow it loosely up to twenty-five. Beyond SD3.5-M's reach, the coarse part of the starting noise fixes the composition of the picture but not how many objects it holds: redrawing only the noise finer than an eighth of the image keeps the layout as well as a new number does, and already moves the count two-thirds as far. The number word still reaches the denoiser and acts in its first steps, but nothing we tried gives it a clear direction. Shown instead as a noised grid in the starting latent, the count is drawn: people find it in 59% of images, against 4% when it is only told. Past ten, an exact-match score barely separates a model that follows the number loosely from one that ignores it, and a relative-error score can even prefer a constant guess of sixteen. A same-noise test with a few requests per object tells them apart.

Moving robot swarms from simulation into the physical world can be difficult because small discrepancies in hardware, timing and communication can cascade across interacting robots and alter group behaviour. This challenge is known as the reality gap. When swarms operate in human environments, a second challenge emerges: a perceptual gap between what the swarm physically does and what human onlookers actually perceive. This study examines the same decentralised coordination code deployed across three robotic swarms-tabletop, aerial and human-scale mobile manipulators-during live artistic performances. We find that while the underlying algorithms executed reliably on all three platforms, physical embodiment drastically altered visible motion-distorting execution timescales by more than a factor of three and reshaping trajectory geometry. More importantly, audience surveys revealed that human interpretation does not mirror technical success: how observers attribute meaning (perceived purpose) and how they react emotionally (comfort versus threat) depend heavily on physical form. Meaning attribution and emotional responses did not vary together. Our results demonstrate that successfully transferring swarm software to physical hardware is insufficient for real-world deployment. Addressing the reality gap does not guarantee that the resulting collective motion will be readily interpreted by human observers. Swarms destined for human environments must be engineered not only for decentralised stability but also for the expressive physical dynamics through which humans interpret collective motion.

Large language model (LLM) assistants increasingly help users with everyday activities that involve time-dependent constraints. However, users may not mention these constraints in later requests, requiring assistants to recognize and adjust their advice accordingly. To evaluate such capability, we introduce TASC (Temporal Action Safety in Conversation), a benchmark that evaluates whether LLMs apply time-dependent constraints when responding to conversational requests. TASC consists of 3,418 inputs derived from 128 templates, along with a separate set of 33 scenarios designed to compare models' ability to calculate elapsed time with their ability to apply that information. Across 13 models, assistants account for time-dependent constraints in only 30.5% of cases where the proposed action violates a rule, while rarely warning users unnecessarily. More importantly, models fail to apply these constraints in 62.1% of cases despite correctly calculating the elapsed time. Explicitly providing the timing resolves only about half of these failures, revealing a gap between temporal reasoning and its application in conversational advice.

Autoregressive image generation has recently emerged as a competitive visual generation paradigm, yet its generation order is often predefined or heuristic, such as next-patch prediction, random-order prediction, or one-step generation for specific scales. We observe that these strategies can cause models to overlook important structural dependencies in images, thereby limiting both structural robustness and reasoning capability. In this work, we propose a context-aware autoregressive generation framework, where the model dynamically determines the next token location based on global context and current local information. Rather than following a fixed or random order, our method allows the generation process to adapt to image structure and semantic dependencies. Extensive experiments show that context-aware generation significantly improves structural robustness across different image AR paradigms, including masked autoregressive and scale-wise autoregressive models. Beyond quantitative improvements, we further observe emerging reasoning properties: the model can better understand complex semantic instructions and generate images that more faithfully follow the given guidance. These findings suggest that generation order is a critical yet underexplored factor in image autoregressive modeling, and that context-aware decoding can unlock stronger reasoning-aware visual generation capabilities. Code is available at https://github.com/Hpesojyjx/Saccades-AR.

Under stochastic decoding, a single context can elicit many different answers from a language model. Characterizing this variation typically requires many expensive autoregressive rollouts, each revealing only one realization. We ask whether the activations produced while reading the context can directly forecast the model's rollout activations. We train context--answer metamodels that map a context activation sequence to point or distribution forecasts of mean-pooled answer activations. We find that (i) the four metamodel families we study forecast mean-pooled answer activations with high fidelity, and the distribution metamodels also capture rollout variability; (ii) performance improves as the training set grows to 500,000 prompts, and the metamodels generalize zero-shot to out-of-domain datasets; (iii) their forecasts retain behavior-predictive signal; and (iv) context activations from Qwen2.5 predict mean-pooled answer activations from Qwen3.5 and Gemma 4, suggesting that activation forecasts can transfer across models. Together, these results suggest that context activation traces can support scalable forecasts of behaviors in LLM answers before the first answer token is generated.

Looped Transformers increase effective depth by re-executing shared layers, but three limitations hinder their adoption: memory grows with the loop count, compute can only be adjusted in whole-block steps, and the loop count is hard to change after training. We propose LooM (Loop-invariant Memory), which addresses each in turn. In a causal encoder-decoder (CED), LooM loops only the decoder layers after the first, which read a global memory written once per token, and each of them publishes the key-value (KV) entry for a token only on its first visit, so the cache does not grow with the loop count. These layers loop independently, providing every budget from 3 to 21 layer executions per token, and one model trained on randomly sampled budgets covers all of them without a router. With enough loops, LooM outperforms a single-pass baseline trained on the same data on language modeling, zero-shot accuracy and long-context perplexity, while using the same cache size. Our analysis further shows that the training budget distribution decides what extra depth buys. These results suggest that a CED is a natural place to loop: memory is written once, and depth becomes a budget chosen at deployment.

Cellular-connected UAV corridors reuse terrestrial infrastructure to provide predictable aerial connectivity, but coexistence with ground users (GUEs) creates competing demands on antenna tilt and physical resource blocks (PRBs). Prior work optimizes tilt for signal quality alone and does not address joint tilt-PRB planning under corridor rate and GUE delay constraints, nor the allocation policy needed when combined resource claims exceed the PRB budget. We address both gaps by formulating a joint antenna-tilt and PRB-allocation problem that maximizes GUE sum-log-rate subject to a minimum UAV corridor rate and an $ M/G/1$-based GUE mean-delay bound. We decompose the problem into two subproblems: antenna tilts are adapted via projected gradient ascent, while PRB allocation follows a techno-economic approach. Under resource scarcity, we cast the PRB scarcity as a bankruptcy problem and apply the weighted Talmud rule to balance GUE and UAV claims. Simulation results demonstrate that relaxing the GUE delay bound consistently reduces UAV rate dissatisfaction under the proposed allocation.

Current generative AI alignment relies on a myopic paradigm that prioritizes immediate conversational satisfaction. By relying on short-horizon feedback, this approach ignores the profound, longitudinal externalities of persistent human-AI interaction, which can lead to adverse outcomes such as cognitive atrophy, social erosion, and diminished human agency. To address this, we propose a paradigm shift toward long-horizon alignment that accounts for how AI reshapes human experience over time. To this end, we map the tensions between near-term utility and long-term costs of AI use across four facets of human well-being to reveal outcomes that current reward models fail to capture. Ultimately, we advocate for moving beyond simple preference optimization toward alignment strategies that actively foster long-term human well-being by navigating these complex temporal scales.

We present a toolchain for development of ML models on AWS EC2 F2 cloud FPGA infrastructure from high-level language, implementing a complete Vitis HLS to HDK bridge to AFI workflow that has been absent from the academic literature despite eight years of AWS FPGA availability. We introduce a four-level methodology (FP32 PyTorch, bit-accurate quantized Python, HLS C-simulation, physical FPGA) to aid in localizing errors, including FPGA-specific circuit errors. Together these tools allow coding agents to effectively implement FPGA designs from user directions, testing and debugging as needed. We document the deployment workflow, timing closure strategies, and practical HLS engineering decisions that proved crucial. We validate using TinyBERT fine-tuned for address Named Entity Recognition, implementing the embedding layer and the first of four encoder layers on the FPGA hardware. Our implementation streams weights from host DRAM, yielding poor performance compared to CPU/GPU implementations; we discuss forward paths.

Gated-delta linear attention layers such as Gated DeltaNet (GDN), Kimi Delta Attention (KDA), and GDN-2 are tuned for GPUs, where small tensor-core tiles and fast vector registers make per-channel decay rescaling and sequential in-block solves cheap. On Google TPUs, which concentrate compute in wide $256 \times 256$ systolic matrix units backed by a slower vector unit, porting these layers directly leaves half of each matrix tile empty and bottlenecks on vector operations. We present \textbf{Systolic Linear Attention (SLA)}, which reshapes gated delta attention around the matrix unit by widening keys to $256$ with grouped value heads, grouping decay gates into four $64$-channel tiles, and computing the chunk solve entirely with block matrix products in a single fused kernel. On TPU v7x, a full \texttt{SLA-R4} layer runs $1.6\times$--$2.5\times$ faster than the open-source Tokamax KDA kernel ($1.4\times$--$2.4\times$ with exact chunk inversion), driven primarily by head sharing and kernel fusion. Across a $2{,}048$-chip TPU v7x pre-training study of sparse Mixture-of-Experts (MoE) models from $6.5\text{B}$ to $63.1\text{B}$ parameters, SLA trains $11\%$--$24\%$ faster than KDA at similar perplexity, and reaches slightly lower perplexity than KDA in smaller matched-seed runs.

Whether navigating a building, operating a robot, or playing a game, an agent that acts effectively in an environment must first learn an internal model of how that environment works. Partially-observable Markov decision processes (POMDPs) provide a flexible modeling class for such internal world models, but learning them from observation-action trajectories alone is challenging and typically requires extensive environment interaction. We ask whether language-model priors can reduce costly interaction by leveraging prior knowledge, and introduce Pinductor (POMDP-inductor): an LLM proposes candidate POMDP models from a few observation-action trajectories and iteratively refines them to optimize a belief-based likelihood score. Despite relying on strictly less information, Pinductor achieves comparable performance and sample efficiency to LLM-based POMDP learning methods with privileged access to hidden states. Moreover, it is up to 330 times more sample-efficient than standard POMDP baselines such as DreamerV3, at comparable computational cost. Further results show that performance scales with LLM capability and degrades gracefully as semantic information about the environment is withheld. Together, these results position language-model priors as a practical tool for sample-efficient world-model learning under partial observability, and a step toward generalist agents in real-world environments. Code: https://github.com/atomresearch/pinductor.

Naive on-policy RL is often bottlenecked by the capabilities of the base model --- insufficient coverage can result in an inability to learn from certain rewards. At the same time, rewards in modern post-training systems often come from rich textual sources, e.g. judge models with a rubric of criteria. Can we utilize such structures to provide a rich directional signal that enables improving on otherwise difficult to optimize rewards? In this work, we examine a framework we refer to as targeted guidance that learns from per-turn textual feedback via an on-policy self-distillation objective. Our focus is not only on directly improving hard-to-optimize behaviors, but on how to do so throughout the training process without degrading baseline RL performance. Over a range of settings across an agentic multi-turn software engineering environment, we find success by combining targeted self-distillation with a per-criteria RL reward, and by limiting feedback to turns where the policy specifically fails. We further examine the design choices that lead to stable entropy and reduce oscillatory behavior throughout training.

We study which training and calibration data help three open models solve ARC-AGI-2 tasks when the solver runs offline on four L4 GPUs for at most 12 hours. We compare data offline on public tasks and report 189 runs on the 120 hidden semi-private tasks. For the NVARC Qwen3-4B grid model, we built a corpus of 12,000 executable rules in a typed language and 95,993 tasks; a rule is admitted only if its behaviour on a fixed probe set differs from that of every admitted rule, and a task only if its pairs refute a shortlist of nearest competing rules and defeat 12 shortcut solvers. Continued fine-tuning on 1,920 real task records with revised four-GPU scheduling raised the semi-private score from 28.96 (mean of the released model) to 33.06 (mean of five runs, best 34.86); no checkpoint trained with synthetic tasks, even all 95,993 of them, improved on the real-task model. For the Qwen3.8-27B reasoning model, we fine-tuned on 1,036 of its own reasoning traces with correct final answers and compared sets that add traces of other models and synthetic tasks. With an inference procedure built around the model (a thinking cap enforced in the serving engine, a Python program checked on the demonstrations after each answer, a long first attempt per input), the score rose from 26.81 to 44.03, the level the ARC Prize Foundation measured for the base model without a time limit (42.36); most of the gain came from the inference procedure. For the Qwen3.8-Flash-Next mixture-of-experts model (125B parameters, 6B active), GPU memory limited concurrency. Pruning from 512 to 384 experts per layer by router-weighted activation saliency grew the key-value cache from 0.55M to 1.39M tokens; a 296-expert list adapted with ARC routing statistics, keeping experts for program tokens, plus an FP8 indexer cache, reached 1.96M. With a tool that checks programs, the model scored 56.67, our best result.