Neural networks, spiking or not, compute with amplitudes: activation values, spike counts, firing rates, never event times. No measure exists of how much a network uses time. We propose one, the degree of temporal representation: the fraction of a network's four functions (encoding, activation, decoding, learning) whose variables or parameters are temporal, and hypothesise that it predicts performance on temporal tasks. At fixed architecture, accuracy and energy efficiency increase with the degree: a delay network of maximal degree reaches 98.94\% on Google Speech Commands with 175k parameters, above transformers. To test whether networks of different degree really differ, we convert the delay network into artificial and spiking networks with the same inputs and outputs. From this common start, the artificial one learns differently and temporal gradients depend on event times. Hence we obtain a design rule: for temporal data, make event times variables of a network's functions. Spiking networks, converted after training, keep the delay network's accuracy, unreachable by spiking networks trained from scratch.
The Practical Molecular Optimization (PMO) benchmark standardized evaluation in molecular optimization, but it is built on simple property-based oracles that do not capture structure-based drug design. As the field has moved to docking-based objectives, evaluation practice has become unstandardized, and different methods report results on different docking tasks, often without a shared protocol for fair comparison or sample-efficiency constraints. To address this, we introduce Practical Molecular Optimization for Docking (PMO-Dock), a benchmark and protocol for docking-based optimization consisting of 25 tasks covering hit generation, lead optimization, and a new specificity task requiring strong on-target binding while penalizing off-target interactions. The protocol separates model development from final evaluation: validation tasks are used for hyperparameter tuning, and test tasks are held out for final evaluation and are not used during development. Evaluation on the test tasks is done under a strict oracle budget to reflect realistic drug-discovery settings. We benchmark four diverse high-performing methods spanning different optimization paradigms, Saturn (reinforcement learning), GenMol (discrete diffusion), Genetic-guided GFlowNet, and ChemLlama (LLM-based evolutionary search). Our results show no universally best method across tasks, additional analyses of hyperparameter transfer (cross-task and cross-category) and sensitivity show that transfer is often limited and varies by method, motivating method-dependent selection strategies. This benchmark provides a concrete standard for evaluating generalization to unseen tasks and sample efficiency under a limited test-time oracle budget.
Summary-token architectures such as Kwai Summary Attention (KSA) compress distant context into learnable Summary token per chunk, and hybrid variants interleave these layers with full attention to retain exact access to the original text. In such hybrid KSA-Full architecture, the interleaving inserted summary tokens are served as a streaming flow role across different layers. Although effective, however, it also introduce some bottlenecks: (1) the different layers summary compression chunk must be same, therefore hard to reach a flexible KV Cache compression, e.g., different layer set different summary compression chunk window. (2) in the full-attention layers, the summary token introduced additional computation cost, but such layer ignore them and still pay quadratic prefilling, therefore unfriendly to the computation side. Motivated by them, we developed the KSA V2, which not only achieves a more flexible KV cache memory compression but also reduces the computation cost in Full attention layer. Specifically, we found that the summary tokens can serve two roles: the KV memory and sparse indexer.
Autonomous research agents must acquire and reason over evidence within a limited budget. Yet relevant sources can remain uninspected or unused, leaving a gap between retrieval coverage and the evidence that informs research. The challenge is to align fast local judgments with changing evidence needs and feed their findings back into research decisions. We introduce JevGPT, a dual-system architecture for autonomous research that couples fast evidence processing with deliberative research planning. System Two sets evidence objectives and integrates findings, while System One locates support, assesses relevance, and guides further inspection. Independent source judgments run in parallel; selected original passages inform the next research step. Experiments on fixed-corpus question answering and document ranking show that JevGPT achieves leading task performance, with faster completion and lower cost than comparably effective research agents. Evidence analysis shows broader source inspection and greater coverage of supporting documents in the research loop, highlighting efficient evidence processing as a route to more capable autonomous research.
Rubric-based reinforcement learning grades a response against many written criteria but collapses them into one reward, which discards the detail the rubric carries and gives no GRPO reward gradient when a group's responses tie. A frozen teacher that reads the rubric as privileged context provides token-level supervision, but we observe an early plateau in the gains over GRPO. RELAY uses advantages and rewards to select positive-advantage and low-scoring zero-advantage rollouts, and monitors the distillation loss to decide when to retire the teacher. Stage 1 combines selective distillation from a frozen teacher with GRPO; a plateau detector then removes the teacher, and Stage 2 continues with GRPO alone. Calibrated on one model-domain pair and reused unchanged in three others, the plateau detector fires between steps 122 and 171. RELAY scores highest in-domain among the compared methods in all four settings, ahead of GRPO on RaR-Medicine by 4.62 points with Qwen3-1.7B and 6.02 with Qwen3-4B. On Qwen3-1.7B/Medicine, removing the teacher from a shared checkpoint outperforms keeping, weakening or refreshing it. Distilling selected rollouts covering about half the response tokens also outperforms distilling all rollouts at a matched average coefficient.
Cybersecurity evaluations of AI agents have focused largely on offense, yet a working exploit does not show that an agent can make software safe. An attacker needs to exploit only one vulnerability; a defender must find and fix all of them. We introduce CWE-bench, a benchmark of 250 repository audit-and-repair tasks spanning 185 CWE types and eight languages, grouped into ten weakness classes. We release 100 tasks publicly and keep 150 held out. Agents are given a repository to audit and a description of its threat model, but are not told the weakness, where it occurs, or how many vulnerable sites there are. Each patch must block the exploit while preserving original functionality, as checked by a deterministic programmatic verifier. Verifiers were curated with panels of agentic judges and reviewed by hand, so that they reject incomplete repairs without penalizing correct ones. The best agent reaches 67% pass@1 and 79% pass@4. We further conduct analyses showing that (i) agents with similar pass@1 differ widely in retry gains and cost, and (ii) no single agent is best across languages or weakness classes. As the offensive capabilities of agents improve, CWE-bench provides a broad and rigorous measure of whether their defensive capabilities keep pace.
Architecture attribute predictors reduce the cost of neural architecture search by estimating accuracy or latency from a small set of evaluated candidates. Neural architectures are operation-labeled directed acyclic graphs (DAGs), where edge directions determine computational roles and execution dependencies. Existing graph Transformer predictors incorporate topological and directional information in different ways, but primarily focus on specific mechanisms rather than coordinated encoder-wide modeling of these semantics. We introduce encoder-wide directional consistency, a design principle that coordinates the modeling of directed computational semantics across node initialization, global spectral encoding, pairwise interaction, and feature updates. To implement this principle, we propose DTA-GT, a Direction- and Topology-Aware Graph Transformer that integrates direction- and topology-aware node initialization, magnetic Laplacian spectral encoding, direction-dual structural attention, and a Direction-Sensitive MoE-FFN. Across accuracy and latency prediction on NAS-Bench-101 and NAS-Bench-201, DTA-GT consistently outperforms representative RNN-, CNN-, GNN-, Transformer-, and hybrid predictors under limited-budget settings. Ablations and magnetic spectral controls support the contribution of the direction-aware stages and orientation-sensitive spectral information, while mechanistic diagnostics characterize module behavior. These results support encoder-wide directional consistency as an effective design principle for architecture DAG representation.
Game-Creation-Agents (GCA) turns a natural-language request into a runnable Phaser 4 browser game. The pipeline first builds a multimodal game design document that couples mechanics and scene contracts to grounded visual references. It then projects that document into a runtime scaffold, machine-readable contracts, and deterministic tests. Eight specialized agents next carry out production: scenes are implemented serially, assets are produced in parallel, and one reduced manifest joins the two branches into a single build. Build-linked real-UI checks and a code-blind critic then verify that build and decide whether the game may ship. Each blocker returns to the agent that owns the affected artifact for a bounded repair, and the game is released once all the gates pass. We evaluate finished games separately, through their rendered interfaces alone, over prompt-matched tasks run under a common protocol. GCA achieves stronger requirement satisfaction, core mechanics, content depth, and functional visuals than a comparable open pipeline. In a side-by-side human study, raters also score GCA's games higher on every dimension and prefer them overall.
Frontier AI systems are increasingly natively multimodal, jointly pretrained on multiple modalities, such as text, vision, and audio. However, the scientific understanding of the multimodal data mixtures used to pretrain these systems --- how much of each modality to mix, and at what potential interference cost to the others --- is largely absent or gatekept. In this work, we derive data mixing scaling laws for three modalities: text, vision, and audio. Multimodal Large Language Models (MLLMs) increasingly adopt Mixture-of-Experts (MoE) architectures, as MoEs scale efficiently and partially limit cross-modal interference. We therefore use MoEs to conduct 268 experiments ranging from 457M to 8.3B parameters, on up to 150 billion tokens. Our work leads to multiple actionable insights into multimodal data mixing for training MLLMs. First, compute optima are modality-specific: audio modality is model-heavy, favoring scaling parameters over tokens, while text and vision require balanced scaling of model size and data. Second, repeating vision or audio data beyond 2x yields negligible benefits. Finally, cross-modal interactions are highly asymmetric: vision strongly interferes with audio, whereas audio's interference with learning text and vision is negligible. Overall, we provide a principled foundation for understanding the multimodal data mixtures needed to train frontier MLLMs.
Large language models are increasingly used as scientific assistants, yet we still lack reliable ways to evaluate whether they can investigate a system rather than merely answer a prompt. A useful evaluation must let agents choose experiments, interpret finite evidence, and make causal decisions, while preserving objective ground truth. We introduce ACED (Active Causal Evidence and Decision Benchmark), a neuro-symbolic benchmark that separates the readable scientific surface from the formal causal world: natural-language scenarios describe the system, while a hidden causal model defines the data, the effect of every intervention, and the oracle answers. In ACED-Repair, agents diagnose a failing production system: within 15 experiments they must decide which controls to test, read confounded and filtered data, and recommend a fix, which we re-simulate on a hidden true objective. In ACED-Query, agents answer causal questions and make constrained causal decisions in worlds backed by categorical probabilistic graphical models. Acceptance on ACED-Repair ranges from 43% to 0% across 12 LLMs, while strong agents approach ceiling on ACED-Query's qualification questions. Tracing every experiment shows where investigations stop: most open-weight models test a true lever on its own in fewer than half of the worlds, the strongest models test the right levers but often miss the working dose or combination, and agents that do run a working fix almost always recommend it. No model reliably repairs systems with two independent causes or systems that need a subgroup-specific policy. These results suggest that reliable AI co-scientists need explicit support for choosing experiments, keeping track of evidence, and composing interventions. Code and data are available at https://github.com/Rose-STL-Lab/ACED-Bench.
World-action models (WAMs) jointly predict robot actions and future observations, but their video diffusion backbones are costly for edge deployment. We introduce SmolWAM, a compact latent WAM derived from the 2B-parameter Cosmos Policy using single-frame encoding, cached text features with preserved attention sinks, depth pruning, and token compaction. With 12 of 28 transformer blocks and 1,008 tokens, SmolWAM predicts actions and future latents in one rectified-flow step. After 3.6 A100 GPU-hours of training, it achieves 95.5% success over 400 LIBERO episodes at 57ms, 2.3TFLOP, and 2.4GiB per 16-action chunk on an A100: 11x lower latency and 3.2x lower memory than the five-step teacher. Its world head predicts scene motion; joint training improves success by 6.2 percentage points over an action-only ablation.
Memory-augmented assistants retrieve evidence from stored conversations and documents to answer queries. Most embedding-based retrievers compare single-vector representations of queries and memories, even when relevance depends on a small part of the content. Such local evidence can be matched to queries using the contextual token representations already computed by the same embedding models. Our method uses sparse token matching to improve retrieval without additional training. Each query token retrieves a small set of matching tokens across candidate memories. The matching scores are aggregated for each memory and combined with whole-memory similarity to rank memories. An optional low-rank query adapter can be trained with supervision while keeping the encoder and stored memory representations fixed. Our method improves retrieval across eight benchmarks and multiple model families without additional training. Query adaptation further increases the gains to 13.3 and 14.3 percentage points in average Recall@10 and NDCG@10 on Qwen3-Embedding-4B. These retrieval gains translate into higher question-answering accuracy on long-term memory benchmarks.
We present DreamTrue, a multi-view, cross-embodiment robot world model for action-faithful and physically plausible video prediction. Training such a model on existing robot datasets faces two obstacles: imprecise calibration can impair action following, while limited coverage of unsuccessful interactions can bias predictions toward successful outcomes. To improve action following across embodiments, we render action trajectories into image-space conditions and introduce offline geometric calibration to align these conditions with the target videos. To broaden interaction coverage, we introduce counterfactual post-training, modifying recorded action trajectories and generating future videos under a wider range of actions and contact configurations. To provide feedback on these predictions without paired ground-truth futures, we construct a human-annotated video dataset covering robot, object, and interaction defects and use it to train an embodied video reward model. Its scores guide reinforcement-learning post-training toward more physically plausible interaction outcomes. On AgiBot, DreamTrue attains state-of-the-art action following, while reducing the human-assessed interaction defect rate from from 48.12% to 6.25%. Notably, our model ranks first in the world model track of the AgiBot World Challenge 2026. The project page can be found at https://brave-eai.github.io/DreamTrue.
Conditional diffusion models generate diverse, novel, and high-quality samples under prescribed conditions. However, theoretical understanding of their memorization and generalization remains limited, while recent works have characterized these behaviors primarily in unconditional settings. In this work, we analyze a random-feature conditional score model in the high-dimensional proportional limit, deriving asymptotic expressions for training and test losses. By decomposing the test loss, we show that in the overparameterized regime, increasing model width improves prediction of the condition-dependent mean while reducing within-condition prediction variance, a phenomenon we term "malign generalization." We further show that in the infinite-width limit, the model predicts the conditional mean as accurately as ridge regression of the data on the condition, while its within-condition variance vanishes. Finally, analyzing the training loss reveals that more informative conditions lead to memorization of training samples at smaller widths. These theoretical findings are supported by experiments with U-Net architectures on realistic data.
Behavior cloning (BC) in non-Markovian environments is a challenging problem because policies have to reason over contextual information over long horizons. Existing policy architectures rely on recurrent or attention-based mechanisms to capture long-term dependencies. However, recurrent models suffer from hidden-state collapse and gradient instability under backpropagation through time while attention-based models are fundamentally limited by context length. To address these issues, we propose Keyframe Mnemonics, a novel self-supervised method that $\textit{discovers}$ a set of information-critical observations ($\textit{mnemonics}$) by learning an objective from randomly sampled past observations and using it as a reward for keyframe selection. We then train a BC policy that conditions on the discovered keyframes to model the action distribution. Under the simplicity bias assumption, our formulation provides context retention guarantees over an infinite horizon, while maintaining a small set of decision-relevant keyframes in the policy's working memory. We evaluate our method on synthetic memory domains, where mnemonic-conditioned BC policies achieve $100$% success rates (SR) and generalize to horizons orders of magnitude beyond training without performance degradation. Moreover, we evaluate on a memory-intensive robot manipulation benchmark, achieving a $13.9$% average absolute SR improvement over the strongest baseline across $23$ tasks. Finally, we demonstrate generalization with $80$% SR at $20\times$ longer time horizons on a real robot. Code and videos available at https://keyframe-mnemonics.github.io.
Steinke and Zakynthinou(2020) introduces the Conditional Mutual Information (CMI) framework of analyzing the information complexity of learning algorithms based on algorithm-dependent information-theoretic quantities. We study one of these quantities, the evaluated Conditional Mutual Information (eCMI). It has been an interesting question whether the optimal PAC guarantee for VC classes can be recovered from the algorithm-dependent analyses via CMI. And we show that it is possible to recover this guarantee by constructing a learning algorithm whose eCMI is of order O(d) in the realizable case, where d is the VC-dimension of the concept class. Specially, our algorithm is a randomized Majority-of-5 base learners with optimal in-expectation generalization guarantee.
GNN-to-MLP distillation aims to retain the predictive accuracy of a message-passing teacher while deploying a graph-free MLP at inference. Existing methods mainly transfer node-wise predictions or use confidence-based reweighting, but they do not specify where the student should preserve the teacher's graph-induced geometry. We show that this omission leads to two spectral failure modes in the student's representation space. On sparse graphs, the student suffers from spectral underfit, missing high-energy teacher directions induced by aggregation. On dense graphs, it suffers from spectral overfit, retaining spurious directions that the teacher has collapsed through aggregation. Motivated by an energy-weighted teacher-student alignment objective, we propose Graph Geometry-aware MLP (G^2MLP), a training-time distillation framework guided by Ollivier-Ricci curvature. Curvature acts as a local proxy for allocating supervision between prediction-level and representation-level alignment. The deployed model is a standard MLP with no graph access at inference. Across node-classification benchmarks, G^2MLP improves over graph-free distillation baselines in most settings, reduces the teacher-student rank gap in both regimes, and transfers to Graph Transformer teachers and link prediction.
Learning systems adapt quickly within a familiar family of models. The harder step is deciding, from a prediction failure, whether opening a richer family is worth its cost. We cast this as a costly sequential decision and introduce the Structural Revision Environment (SRE), which produces matched failures from noise, exceptions, change within the family or structure outside it, varies the price of expansion and the remaining horizon independently of the evidence, and admits an exact normative solution of the one-shot decision. The solution shows that revision is a value boundary, not an evidence threshold: except at the lowest prices, the richer family becomes the more probable explanation long before expanding becomes optimal. Small Transformers trained in SRE reproduce the expansion boundary from utility alone. Pretrained 7B and 8B language models fail at a definite point: their predictions carry a failure-sensitive signal that their decisions do not reflect. Told the gain of expanding, they weigh it against price and horizon only when they reason, and the one model that finishes its reasoning on the history alone still does not estimate that gain. Controlled post-training of the small Transformers moves the prior and the sharpness of predictions, and neither moves the criterion.
Multimodal large language models (MLLMs) have made significant progress in visual understanding, but precise 3D spatial reasoning integrated with physical environment remains difficult. Furniture assembly requires not only recovering step-level operations from diagrammatic manuals, but also translating semantic attachment relations into 6D pose updates that enable parts to physically interact with the environment and previously assembled components. To study this problem, we propose AssemState, a zero-shot framework for manual and physical-state-guided furniture assembly. It firstly employs anchor-guided boundary assembly states to decompose manual pages into single-part operations and recover an assembly-tree. Then, it uses iterative after-state feedback refinement to guide successive (SE(3)) updates and corrections, and validates their physical plausibility through simulation-based release tests. Experiments show that compared with the strongest prior baseline, AssemState improves F1 from 38.58\% to 62.80\% and Tree Exact Match from 28.24\% to 53.92\% for assembly-tree recovery. On 243 independently evaluated part-level operations, our proposed iterative refinement improves judge-accepted operations from 0 to 5.3\% and reduces mean Chamfer distance from 5.4111 to 1.7744. However, visually plausible candidate poses may still suffer from collision, floating, mirror-orientation errors, incomplete seating, and wrong-side attachment. These results show that AssemState improves operation-structure recovery and selected local pose metrics, while MLLMs remain limited for spatial relationship reasoning.
The logarithmic vector multiplicative error model (log-vMEM) provides a methodology for modeling and forecasting multivariate positive-valued time series. However, the number of autoregressive parameters increases rapidly with the system dimension and lag order, posing challenges for estimation and model selection. This paper develops a regularized estimation framework for the log-vMEM with multivariate Gamma errors by incorporating hierarchical lag structures to induce sparsity and accommodate heterogeneous lag dynamics. We consider three hierarchical lag structures: componentwise, own-other, and elementwise, in combination with group lasso, adaptive group lasso, group MCP, and group SCAD penalties. Parameter estimation is carried out through an alternating block-coordinate optimization algorithm with sequential Gauss-Seidel updates, while GPU-accelerated numerical quadrature is employed to address the computational burden of evaluating the multivariate Gamma likelihood. Simulation experiments investigate parameter recovery and lag selection under different sparsity structures. The real data analysis is carried out on the realized volatility measures for Microsoft (NASDAQ:MSFT) which allows a comparison of the competing regularization approaches in an empirical financial setting. Together, these developments provide a computational framework for sparse multivariate volatility modeling with flexible lag structures and dependent Gamma errors.