Real-time voice agents need a speech recognizer that never waits for future audio and whose memory per call does not grow with the length of the call. We present E-Gaborformer, a strictly causal encoder that combines Gated DeltaNet-2 recurrent layers with sliding-window attention in a 3:1 ratio, and builds every temporal convolution from a fixed dictionary of causal Gabor atoms. Its streaming state is constant, about 19 MB per call at production scale including the prediction network, where a causal-attention encoder of the same size accumulates over 1 GB after ten minutes. In a controlled causal study on LibriSpeech, E-Gaborformer leaves the initial plateau earlier, reaches a given accuracy up to 18% sooner, and matches or beats an E-Branchformer trained with the identical front end and recipe. At production scale, a 0.6B-parameter Hebrew and English streaming transducer built on E-Gaborformer has 28 to 31% lower word error rate (WER) than NVIDIA's Nemotron 3.5 ASR Streaming, a model of the same size, on English voice-agent traffic, and up to 20% lower Hebrew WER than our previous production system. The same weights serve strictly causal and 480 ms-lookahead decoding with no loss in accuracy, and a two-class formulation of the RNN-T lattice trains the transducer 2.6x faster with exact loss and gradients.
Despite recent progress in handwritten text recognition, medieval Arabic scientific manuscripts remain an open challenge for current vision-language models (VLMs) due to cursive ligatures, variable diacritization, and more importantly the lack of training data. This paper aims to address this gap, by proposing the first annotation guidelines dedicated to the transcription of scientific Arabic manuscripts inspired by textual criticism, a well-established approach to historical text editing. We then propose, ATHAAR, a new dataset manually annotated following this guideline. Unlike previous resources in the field, ATHAAR varies both witnesses of the same text, calligraphic styles, and content to deal with texts but also numerals. Relying on state-of-the-art VLMs, we finally design a set of experiments to study models transfer across scripts, chapters, and treatises. Our findings highlight the limitations of current systems when faced with the specificities of ancient scientific texts.
Mixture-of-Experts (MoE) models route each token to a few of many experts, yet what an individual expert is responsible for remains unsettled. Existing analyses count how often an expert is selected on domain-labeled corpora, an instrument that fixes the taxonomy in advance, stays coarse, and records selection rather than contribution. We instead study specialization in feature space by training a sparse autoencoder (SAE) on the output of an MoE layer. Since this output is a router-weighted sum of expert outputs and the SAE encoder is affine, the pre-activation of every latent decomposes exactly over the experts. Aggregated over a corpus, this gives each expert a dominant set of latents whose activation it principally drives. On layers 4, 30 and 46 of Qwen3-30B-A3B, contribution is far more concentrated than gating, most latents have only one or two dominators, and the overlap between dominant sets concentrates on a few expert pairs. Auto-interpreting these sets reveals a division of labor below coarse domains, with distinct experts for Java and front-end web code, or for geometry and for arithmetic and polynomials, while the coarse domains themselves take very unequal numbers of experts. Our SAEs and code are available at https://github.com/Bytedance-Seed/SAE-Interprets-MoE.
Graph Attention Networks (GATs) are increasingly used in domains where scientists need to read a model's reasoning, from medicine to materials science, yet the routing they learn remains opaque to most explainers. Layer-wise Relevance Propagation (LRP) follows the model's computation, but in an attention layer the relevance reaching a node is the sum of two mechanically different quantities: its content relevance through the value path and its routing relevance through the attention path. Existing rules either drop one or add them into a single number. We write a single product rule with free weights on the two factors that recovers current approaches as special cases. The value path measures what content a node supplied against a zero baseline; the attention path measures how the node shaped what was passed relative to its neighborhood, and its sign records whether the node's features moved the gating the way the prediction demanded. On a ground-truth classification task, we show that the same node receives negative, positive and near-zero relevance under three settings of the rule. The separated relevance maps yield a hypothesis about the model: it gains by not reading that node. We give reading rules for both paths and argue that they should be interpreted separately.
Human cognition organizes continuous experience into bounded units while maintaining continuity across them. Linguistic elements are also organized into bounded structural units such as words and phrases, with sentence marking the final bound of a grammatical construction process. How do such boundaries emerge? We propose boundaries emerge from interacting demands to maintain unresolved dependencies, integrate completed relations, and update semantic context, with structural completion allowing but not determining closure. First, in a generative simulation without predefined boundary positions or constituent labels, these demands produced sentence-like segment-length distributions and distinct characteristic timescales, with structural completion allowing but not determining closure. Near boundaries, unresolved dependency load declined, integration cost peaked, and accumulated pressure reset. Then, we tested whether same demands predicted human listeners' boundary choices during comprehension. Listeners were more likely to mark boundaries with fewer unresolved dependencies, higher integration cost and semantic mismatch, and elevated preceding memory load, beyond syntactic and prosodic cues. Third, in naturalistic self-paced reading, sentence-final slowing coincided with declining memory load and increasing integration cost and semantic mismatch, broadly consistent with the simulation. Processing same narratives with a large language model revealed increased local intrinsic dimensionality and straighter hidden-state trajectories at sentence endings, followed by larger representational updates. Boundary explained additional variation in reading time and hidden-state geometry beyond processing demands, but considerably less in next-token surprisal. Together, these findings suggest linguistic boundaries emerge as discrete manifestations of transitions in continuous processing dynamics.
Efficient deployment of large language models (LLMs) on mobile devices requires balancing performance with compute and memory constraints. Post-training compression is an effective approach to managing this trade-off, while requiring much less computation than in-training counterparts. We propose DictLLM, a dictionary-based compression scheme in which weight matrices within grouped layers are decomposed into a shared dictionary and layer-specific sparse coefficients, directly obtained from a pre-trained model without further training. Specifically, we cast DictLLM as a four-step post-training optimization. First, we obtain shared dictionaries via singular value decomposition (SVD) of input-scaled weight matrices, amortizing rank across grouped layers; second, we sparsify the layer-specific coefficients to the target compression ratio using a dictionary-aware Hessian; and third, we refine the dictionaries given the resulting sparse coefficients. Finally, we aggressively quantize both components using factor-specific Hessian information: input covariance for the dictionaries and the dictionary-aware Hessian for the coefficients. The resulting representation reduces memory footprint by storing a single shared dictionary per layer group, along with sparse layer-specific coefficients. On language modeling and zero-shot reasoning benchmarks, DictLLM consistently outperforms existing post-training low-rank compression methods and pushes the Pareto frontier between model size and performance. DictLLM compresses the original LLaMA-7B model by 5.1x with only a 2.9-point increase in WikiText-2 perplexity, achieving a 3.6x higher compression ratio than the state-of-the-art low-rank method.
Scientific figures give research ideas a visual form through choices of abstraction, composition, and emphasis. Generative AI offers new ways to realize these choices, raising questions about explanatory clarity and purposeful visual expression. We examine these questions through a visual retrospective of 371 motivation and framework figures from 200 papers. The corpus comprises a quota-based main sample of 160 papers spanning four temporal periods and a separate purposive style-contrast sample of 40 papers. A single multimodal judge evaluates six dimensions of communication and design, with three repeated judgments per figure. We attempt source-constrained revisions of all 371 figures, of which 369 pass the recorded fidelity checks. Original-figure ratings show no uniform decline across periods. Revisions increase mean clarity ratings by 0.428 points and decrease aesthetic coherence ratings by 0.111 points on a five-point scale. Gains in the six-dimension average are larger for main-sample motivation figures than for frameworks. Case analyses show that added explanation can clarify comparisons while leaving typographic hierarchy and visual economy unresolved. These model-based findings support researcher-led design, with AI assisting exploration and authors retaining responsibility for explanatory purpose, composition, and scientific relationships. The accompanying atlas and evaluation records are available at https://ntu-ai4x.github.io/Rethinking_Figures/index_en.html.
Grounded reasoning, where a model outputs text coordinates of relevant locations in chain-of-thought reasoning, is a common strategy for solving complex visual reasoning problems. However, the benefits and limitations of grounded reasoning are understudied. We study where and why grounded reasoning is most useful. To do so, we introduce a chain-of-thought analysis method to measure how much a model uses its outputted grounding in its thinking trace and perform a detailed ablation study of a grounded reasoning training recipe. We find that grounded reasoning improves counting but shows no significant gain over baselines without coordinate grounding across a wide range of spatial reasoning and perception benchmarks. Using several synthetic tasks, we show that grounded reasoning is most useful when a task requires undescribable references, which are visual concepts that are difficult to describe in language. However, we observe that many existing visual reasoning benchmarks are solvable with only describable text references. To close this gap, we introduce BrickVision, a new benchmark that requires undescribable visual references to solve.
LLM-guided evolutionary search uses AI agents to improve programs for mathematical and computational problems by iteratively refining solutions and combining high-scoring methods. However, the process is often slow because the LLM generates a long textual chain of thought for each round and must repeat this generation many times to maintain population diversity. We present SwarmEvolve, a framework in which hundreds of low-cost agents per round work in swarms, reason in latent space without decoding to text, and share information by passing KV caches. Each swarm studies the problem from a different angle. A novelty filter keeps the most distinct agent ideas by directly measuring the distances between the agents' hidden state vectors. Crossover of ideas happens naturally in latent space, as a summarizer agent receives the combined KV caches of all swarm members. On 21 problems in mathematical discovery and scientific computing, SwarmEvolve gets the highest average score and needs a median of only 3 to 8 minutes to reach the best scores that other state-of-the-art frameworks reach in 60 minutes. On several problems, SwarmEvolve with Qwen3 8B improves on the solutions that AlphaEvolve first found with frontier LLMs. Our work shows that agent swarms that communicate in latent space offer a fast, low-cost approach to program search for scientific problems.
AI-Scientists arrived, but a world for them to grow in did not. Human scientists take theirs for granted $-$ centuries of open problems and shared ideas, waiting in the scientific literature. To AI-Scientists, that world is readable but not runnable, so we build Scientist World for them: every research paper becomes an executable environment. Our autonomous pipeline extracts each paper's open-ended research question and reproduces its proposed method as the baseline; an AI-Scientist entering the environment is evaluated against what the human authors achieved. We release the first 1,000 verified environments. With the world as an evaluation, we found that frontier AI-Scientists can already surpass the human researchers 74% of the time. As a record of how AI does science, it contains 6,400 hours of research traces, with every idea, experiment, failure, and success. As a training ground, we are able to raise (+158%) a small model to match frontier models in research ideation. At last, AI-Scientists' own world has arrived, grown from and growing with science itself.
ML code should run where researchers choose, not where it was first written. Today, machine learning code mainly lives in two ecosystems, PyTorch on NVIDIA GPUs and JAX on Google TPUs. Migrating repositories between them means building and testing across frameworks and hardware. This migration is (1) a practical need, since it unlocks different hardware for researchers, and (2) a natural benchmark: it involves different frameworks and hardware, covers whole repositories, and is long-horizon and challenging. We introduce XStack, a benchmark that evaluates agents on migrating ML repositories between these two ecosystems. Each agent receives a PyTorch repository and a clean TPU machine and must deliver a JAX repository. Its output passes if it meets our grading criteria, based on the source repository's published result, authenticated and calibrated by us. Migration by agents remains unreliable: Claude Opus 5 passes 75%, GPT-5.6-sol 45%, and every other agent at most 40%. XStack is a step toward freeing ML code from its ecosystem. We release XStack to make progress toward reliable migration across frameworks and hardware measurable.
Interpretability, AI auditing and assessments of machine consciousness increasingly rest on mechanism-level quantities: functionals of a system's causal transition law that describe how it computes, not what it outputs. Experiments see that law only through finite, partial and noisy interventions. We ask when such a quantity is identifiable from an experiment's access, continuous in the mechanism and certifiable from finite data, and answer exactly for a deliberately demanding case - the integrated information $\varphi$ of integrated information theory - in stochastic networks of four and five units with known ground truth. Behavior can be silent: a recurrent and a feed-forward network with identical input-output laws at their response delay carry 0.28 and 0 bits, yet internal recordings separate them within 4-53 trials. Partial access can be just as blind: in our examples, units left out of a one-step recording impose precision floors of 0.30-1.02 bits that no sample size lowers, and in all-but-self parity networks, recordings of any length see several hidden units only through the product of their fidelities - a symmetry that one intervention broke in exact checks. The measure has sharp edges: because $\varphi$ selects a partition, and the cause and effect it specifies, by one criterion and reports another, infinitesimal changes of mechanism that cross an edge change it by 25-100% at 99% of the edges we found in random networks, and about 1% of surrogate models that fit interventional data almost perfectly misreport it by more than 20%. One-sided claims survive: the same data certify a lower bound on both sides of an edge, and we prove $\varphi>1.004$ bits throughout a 64-dimensional region of mechanisms and certify it from about $10^7$ controlled transitions with false-certificate probability at most $α$. Measurability is thus a design question with computable answers.
To interpret a direction in a language model's residual stream, we can look for text that strongly activates it. But when such text is found by scanning a corpus, it is limited to what that corpus contains. We develop Maximally-Activating Example Metamodels (MAEMs), language models which map residual stream directions from a subject model to text exemplars that strongly activate those directions. Our objective is to maximise the cosine similarity of a supplied residual stream direction with the activations of the text generated by the MAEM. This label-free RL objective helps to scale MAEMs on diverse sources of residual stream directions. Empirically, we find the resultant metamodel produces informative exemplars: a Qwen3.6-27B MAEM inverts held-out activations better than corpus search or a natural-language autoencoder (NLA), beats corpus search (at 10M tokens) on a majority of held-out languages/domains, and activates 89% of held-out SAE features. It also identifies concepts behind steering vectors about as well as NLAs, recovers the payloads of LoRA backdoors more often than corpus search, and can even verbalise some model workspace inner monologue. Unlike explanations produced by other metamodels (e.g. NLAs), MAEM exemplars can be verified with a forward pass on the subject model, making the MAEM objective a grounded alternative for scaling the training of metamodels.
Adversarial Training (AT) is an important method for improving model robustness. We study the dynamics of deterministic Gradient Descent (GD) in AT. Most existing works in this direction focus on the small or vanishing step sizes regime, where GD resembles a continuous-time flow. However, such step sizes can make training inefficient and are rarely used in practice. In this paper, we study the AT dynamics of GD with large constant step sizes. We first study GD in adversarial linear logistic regression, focusing on two aspects: where the model parameters converge (implicit bias) and how quickly the losses decrease (optimization). We then turn to neural networks training and study how the sharpness of losses evolve during AT. Our results advance the understanding of the behaviors of GD with large step sizes in AT and highlight three key differences from its behavior in Standard Training (ST): (i) the implicit bias gap between AT and ST, previously established only for small step sizes, persists for arbitrarily large constant step sizes, (ii) different from ST, there is no uniform rate of convergence for the optimization error in AT, and (iii) sharpness can exhibit two-stage dynamics that differ from the Edge-of-Stability behavior observed in ST.
On-policy distillation (OPD) trains a student on its own rollouts under a frozen teacher's per-token supervision. Does this dense signal reduce the training FLOPs needed to reach a target accuracy compared with reinforcement learning? We systematically compare vanilla reverse-KL OPD with Group Relative Policy Optimization (GRPO) across task difficulty, student and teacher size, domain, and optimization choices, measuring held-out accuracy against training FLOPs. With an external post-trained teacher, OPD generally loses to GRPO in accuracy per training FLOP at matched settings, even when the teacher is further trained with GRPO, and often suffers response-length inflation. Yet a different teacher changes the picture: when the teacher is the student's own GRPO checkpoint, OPD recovers that checkpoint's accuracy on math with $50\times$ fewer distillation FLOPs than GRPO needs to reach the same accuracy. Following the same teacher setting, we consistently observe this high efficiency with a larger student, on code, and in a second model family. On a single domain, however, the GRPO checkpoint is already the desired model, so distilling it adds to total compute. Across domains, separate GRPO checkpoints hold different capabilities, and no single checkpoint serves every domain. Multi-teacher OPD (DeepSeek-AI, 2026) consolidates per-domain GRPO experts into a single student that matches each expert's accuracy, with the distillation stage costing only 5.2% (on 2 domains) and 8.0% (on 4 domains) of the specialist training FLOPs. Counting specialist training and distillation together, this pipeline reaches joint GRPO's performance on the 2-domain mix with 42.0% fewer total FLOPs.
Models for physical AI tasks, such as robotics and autonomous driving, need not reason solely in language. They can use visual plans--points, bounding boxes, and trajectories--to represent spatial goals and anticipate motion. However, autoregressive vision-language models generate these plans one coordinate token at a time, making richer predictions increasingly costly. We introduce VFlash, a method for fast, accurate visual planning through Flash Primitive Decoding. VFlash exploits the known structure of visual primitives to predict their coordinates jointly, using a lightweight drafter that generates an entire primitive in one forward pass. The layout adapts to the primitive's type, dimensionality, and length, supporting compact boxes and long, variable-length trajectories. Crucially, spatial accuracy need not require reproducing an autoregressive model's exact token sequence. VFlash therefore trains on annotated primitives and directly accepts structurally valid predictions without target-token agreement. Across the evaluated 2D and 3D robot trajectory tasks, VFlash consistently reduces prediction errors relative to autoregressive decoding. Preliminary single-rollout measurements indicate approximately 25x faster generation for 3D point motion forecasting. A world action model conditioned on VFlash plans achieves 70.2% average success across four MolmoSpaces task suites, compared with 62.6% for the baseline.
A defender anticipating an attacker needs a probability distribution over the attacker's next action from information available before that action occurs. Two structures complicate this: the attacker's behavioral regime persists and can switch within an episode, and each action combines an operation with state-dependent arguments whose feasible values change over time. We capture both with a hidden Markov model whose latent state is the active regime and whose regime-conditioned action law is defined on a feasibility-masked tree of executable actions. The same likelihood forms the next-action distribution and updates the regime belief once the action is seen, and a partially observed prefix contributes its resolved local factors. Estimation separates into convex logistic fits that Baum-Welch reuses when regimes are unlabeled. We derive a reach-weighted nodewise KL expression for regime information, a sequence-wise log-loss bound for switching regime paths, and a uniform fallback guarantee. On the CAGE 2 cyber defense benchmark the filtered latent state is worth 0.8 nats per step, and after a controller switch the loss begins to fall in three steps against ten for a recurrent baseline.
Cyclist fatality rates remain high despite policy support for active travel, and autonomous vehicles add complexity to mixed traffic. Neither crash statistics nor multi-model detection pipelines scale to continuous monitoring of perceived safety along a route. We present an edge-assisted crowdsensing system: handlebar cameras supply street-level frames, a vision-language model (VLM) turns each frame into structured traffic attributes through a single prompt, and an explainable boosting machine (EBM) combines them with OpenStreetMap road infrastructure and on-bike kinematics to predict a crowdsourced 1-9 safety score. Extraction belongs at the network edge: the fastest self-hosted model needs 2.49 s per frame and scoring takes 0.4 ms on a CPU. On 1,362 annotated frames from two London routes, the EBM regressor reaches 72.2 percent binary accuracy and a mean absolute error of 1.011 under five-fold cross-validation. It is within one accuracy point of black-box ensembles. Its per-feature shape functions give exact attribution: physical lane separation is worth +0.97 rating points, a painted lane +0.70 and a construction site -0.65. Planners can therefore price an infrastructure change in rating points, and the results enter cost-benefit analysis directly. We also report inter-annotator agreement (alpha = 0.4695), an attribute audit, and a sensitivity analysis linking attribute error to accuracy. With re-encoding, a 20 s sampling interval serves 1,000 riders at 54 Mbit/s at a cost of 1.4 accuracy points, making continuous monitoring practical.
We introduce SM4RT, a multi-task cascade feed-forward model for dynamic scene reconstruction from RGB videos. While existing feed-forward approaches typically address geometry reconstruction and motion estimation separately--and often predict motions in a single coordinate frame that must account for camera movement--SM4RT solves these tasks sequentially by first reconstructing 3D geometry and then bootstrapping motion estimation from the predicted structure. By expressing each pixel's 3D motion in its corresponding camera frame, SM4RT reduces ambiguity arising from camera motion. To address the sparsity of real-world motion labels, SM4RT further incorporates a weakly supervised loss that regularizes unlabeled point motions using labeled point positions at the other time step. We evaluate motion estimation on three dynamic-scene datasets--Kubric, Waymo, and Point Odyssey--and SM4RT achieves the best performance on most metrics compared with prior baselines. We additionally benchmark geometry reconstruction on five dynamic scene datasets and show that SM4RT achieves competitive performance with state-of-the-art methods. Overall, our unified framework demonstrates that combining geometry reconstruction and motion estimation in a cascade not only avoids performance trade-offs but also enhances motion estimation through improved geometric reasoning.
LLM-based medical assistants that follow a patient across visits need a memory that keeps the patient's clinical state correct as it changes. We examine the errors of five memory systems on longitudinal medical dialogue and find three recurring causes: the retrieved evidence gives an outdated state, includes another person's health facts, or lacks a safety fact the answer depends on. We present MedGraph, a graph memory built for these three cases. Construction extracts dated facts from each session, attributes each fact to the person it describes, and links medication, diagnosis, and measurement records into chains that keep earlier states. Retrieval ranks the patient's chains as units, reads each chain at the time the question asks about, and adds the original dialogue text; a patient profile lists every recorded safety fact with every question. On MedMemoryBench, MedGraph has the highest overall accuracy with both a GPT and a Qwen backbone, 10.5 and 10.4 points above the strongest of five baselines, with efficient memory construction and answering in the GPT setting.