Papers

Embedding structured objects into Euclidean spaces has enabled a wide range of successful machine learning applications. Such objects include words, documents, image patches, time series, and graph nodes. In contrast, embedding entire graphs remains a challenging problem. Existing methods either sustain the original order of the graph nodes or match the output nodes to the input ones, both of which create scalability issues. In this work, we propose a graph representation as a cloud of hyperballs, which allows us to define a specific, typically unique, node ordering. Based on this representation, we propose GeoGAE, an autoencoder, in which the Transformer encoder translates a hyperball cloud into a graph-level embedding, and the Transformer decoder translates the graph-level embedding back into the graph. This formulation enables the model to capture both the global graph structure and local relational patterns. We evaluate our method on multiple graph datasets, spanning various domains. The results demonstrate effectiveness of our method in encoding and reconstructing graphs from their embeddings.

Principled exploration in reinforcement learning requires an agent to quantify its epistemic uncertainty and act to resolve it. Uncertainty over the value function provides a natural signal for exploration, yet existing deep approximations remain brittle and perform inconsistently. The central challenge is therefore to scale these ideas robustly. We conduct a systematic empirical study of how epistemic uncertainty is represented, propagated, and optimized in deep epistemic value functions, and uncover distinct failure modes along each of these axes. These findings motivate DEVOTE, a model-free reinforcement learning algorithm that controls how uncertainty generalizes beyond observed data, stabilizes its temporal propagation, and preserves adaptation to the resulting non-stationary exploration objective. Across reward-free exploration and challenging continuous-control tasks, DEVOTE reaches novel states more effectively and achieves higher task return than strong model-free and model-based exploration baselines. These results provide evidence that deep epistemic value functions are a promising path toward scalable, principled exploration.

Sparse autoencoders (SAEs) expose features that help us understand and steer language models, but faithful reconstruction does not guarantee informative concepts. Token-level objectives reward lexical and formatting details alongside semantic content, all competing for a limited sparse budget. We introduce a family of chunk-level SAEs that encode mean-pooled activations over chunks, each a contiguous span of tokens: Mean-Chunk reconstructs the observed chunk, Cross-Chunk predicts an independently processed neighbor, and Joint-Chunk combines both targets. These designs separate the effect of a larger observation unit from that of predicting information shared across passages. With matched training data, chunk-level SAEs remain powerful interpretability tools while learning reliable semantic features that capture high-level concepts and respond selectively to relevant content. Their strengths are complementary: Mean-Chunk improves high-level feature discovery, reasoning detection beyond surface cues, and steering; Cross-Chunk leads document retrieval and classification transfer while producing selective, persistent features. Changing what an SAE sees and predicts yields reliable semantic features for more meaningful tasks. We demonstrate their practical value through gains across downstream tasks such as retrieval, reasoning detection, and steering.

On-policy distillation (OPD) trains a student language model with dense feedback from a stronger teacher on student-generated trajectories. Yet standard OPD weights token-level distillation terms uniformly, implicitly treating local teacher preference as a proxy for correction utility. A decision's task value, however, depends on how the student completes the subsequent reasoning. This mismatch can cause imitation to suppress viable student strategies or reinforce paths the student cannot reliably execute. Verified trajectory outcomes provide complementary evidence about continuation quality, but do not directly identify the utility of individual decisions. We introduce Reward-Aligned Reweighting for On-Policy Distillation (R$^{2}$-OPD), which uses outcome agreement and the magnitude of teacher–student disagreement to continuously reallocate teacher supervision. It gives reward-aligned corrections greater relative influence while retaining dense feedback, moving beyond uniform imitation and hard filtering. Our analysis formalizes the mismatch between local teacher preference and student continuation value and establishes sufficient conditions for reallocation to improve first-order task progress over uniform OPD. Across seven mathematical reasoning benchmarks, R$^{2}$-OPD achieves the highest average accuracy among the compared training methods in both cross-size and same-size distillation. It outperforms standard OPD on all seven benchmarks, with average gains of 3.5 and 2.4 percentage points for 1.7B and 4B students, respectively. An extension to code generation yields an average gain of 1.6 percentage points over standard OPD. These results highlight outcome-guided supervision allocation as an effective way to translate dense teacher feedback into stronger student performance across model scales and task domains.

Scientific discovery requires not only recovering mathematical laws that describe observable behavior, but also identifying the mechanisms that generate them. Existing benchmarks for symbolic regression and scientific agents primarily evaluate phenomenal-law recovery, leaving mechanism discovery largely untested. We introduce MechBench, a benchmark that explicitly separates these two capabilities. Each task is defined by a mechanistic model, a structured set of scientifically meaningful relations whose joint consequences entail an observable phenomenal law, while agents receive only observational data and scientific context. We evaluate mechanism recovery through mechanism probes, which query internal scientific consequences that cannot be inferred from the phenomenal law alone. To reduce reliance on memorized textbook mechanisms, we construct unfamiliar variants through controlled, scientifically interpretable mutations of canonical mechanisms, and screen for mechanistic indistinguishability to exclude ambiguous instances admitting comparable competing mechanisms. Experiments across representative scientific agents reveal a substantial phenomenal–mechanism recovery gap: for Codex with GPT-5.6-sol, phenomenal-law accuracy reaches 35.00% on the Core-set while mechanism accuracy is only 13.75%, with mechanism recovery failing in 64.29% of cases where the phenomenal law is correctly recovered. The gap widens as mechanisms become increasingly mutated, and even providing the correct phenomenal law leaves mechanism recovery below 50%. These results reveal a substantial generalization gap in mechanistic reasoning and establish mechanism discovery as a distinct challenge beyond recovering observable scientific laws.

In ecology, psychometrics, and the analysis of social and financial networks, binary matrices are often analyzed conditional on their observed row and column sums, which restricts the problem to a finite sample space of matrices with the same margins. Two fundamental problems are to count this space and to sample uniformly from it. Sequential importance sampling (SIS) addresses both with independent weighted samples and an unbiased count estimator, but its efficiency depends critically on the proposal distribution. Existing proposals are analytically designed, and their accuracy can vary substantially with the margins. We show that the ideal SIS proposal, under which every weight equals the count and the variance vanishes, is exactly the policy of a generative flow network (GFlowNet) with unit reward on every matrix that has the given margins. We therefore propose MarginFlow, a framework that turns the design of the proposal into a learning problem and amortizes it across margins by exploiting their self-similarity. Every partial matrix is itself an instance with reduced margins, so one set transformer that reads the remaining margins serves every margin. We train MarginFlow on a pool of 1904 margins and evaluate it zero-shot on 1190 held-out margins, synthetic and real, from $3\times3$ to $870\times6$. On 1187 of the 1190 margins it matches or beats the best of 31 analytically designed configurations, chosen post hoc for each margin, and its median effective sample fraction is 99.8%. On the 56 margins where that best loses more than one nat of effective sample size, MarginFlow wins every one and raises the median effective sample fraction from 10.3% to 94.1%.

Self-training generative models - the continued improvement of a model using its own outputs - is becoming increasingly important as high-quality training data becomes scarce. However, naively finetuning on model-generated samples leads to degradation through model collapse and the model autophagy disorder. Negative-guidance self-training methods turn this degradation into a useful signal, using a model finetuned on its own outputs to guide the original model toward improved generation. Existing methods, however, take the negative signal in standard model outputs as given. We instead ask whether this signal can be explicitly strengthened. We introduce Geometrically Modified Outputs (GMOs), which reweight the singular values of the generator's input-output Jacobian to increase the influence of its leading singular directions. This geometric modification amplifies the mode-seeking behavior and distortions of standard outputs, providing a stronger and more targeted negative signal for self-training. Across a range of one-step generative models, GMOs consistently improve the performance of negative-guidance methods, including Neon and SIMS, compared with using standard model outputs.

We study on-policy distillation (OPD) through the lens of reinforcement learning, establishing a connection between the reverse-KL objective in OPD and KL-regularized policy optimization. Building on this connection, we introduce Least-Square Policy Distillation (LSPD), an RL-inspired framework that brings optimistic exploration and off-policy data reuse from value-based RL into policy distillation. LSPD preserves policy diversity through exploration while improving rollout efficiency by repeatedly learning from previously collected trajectories. Our theoretical analysis connects LSPD to optimistic value-based learning and shows that its idealized formulation achieves a sharp $\tilde{\mathcal O}(\log K)$ regret bound under online exploration. Empirically, LSPD consistently outperforms existing distillation baselines across six mathematical reasoning benchmarks and diverse teacher-student settings, with average gains of +1.59 points in Avg@16. Remarkably, through Pass@k evaluations up to k=64, we found that LSPD better preserves policy diversity by achieving stronger performance as k grows. Its fully off-policy variant achieves comparable performance to vanilla OPD using only the first 25% of rollout batches. Together, these results provide an RL perspective on OPD that offers both a principled interpretation and a practical route toward more effective and rollout-efficient language model distillation.

Multimodal prediction relies on diverse forms of evidence: information repeated across modalities, cues specific to a single source, and complex cross-modal dependencies that emerge only when inputs are considered together. While recent methods promote richer interactions, they lack a principled way to isolate these target-relative contributions within learned continuous representations. We introduce a framework that applies contrastive or masked objectives at intermediate layers, coupled with source-wise invertible normalizing flows and a supervised, low-rank latent variable model. This architecture explicitly factorizes the joint distribution into shared task-relevant variation, modality-specific predictive variation, and task-irrelevant dependence. Drawing connections to prior multimodal learning assumptions, our approach evaluates how modalities independently and jointly contribute to the target. Ultimately, this framework unites intermediate representation learning with structured likelihood-based guidance, offering a practical latent-variable lens for characterizing continuous multimodal interactions. Empirically, we demonstrate the effectiveness of our approach across diverse multimodal benchmarks, showing robust improvements in predictive performance.

Recent agentic symbolic regression approaches increasingly rely on large language models to analyze data, select scientific operations, and refine hypotheses over long search trajectories. In such systems, performance depends not only on the underlying model and search strategy, but also on the runtime infrastructure that supports scientific search. We introduce SRHarness, a domain-specific harness for agentic symbolic regression built around three mechanisms: composable scientific actions that provide a common interface over raw, transformed, and candidate-derived quantities; persistent scientific state that retains evaluated hypotheses and exposes compact model-facing views; and trajectory lifecycle management that coordinates continuation, branching, restart, and termination. On LLM-SRBench, SRHarness consistently improves both numerical generalization and symbolic recovery under matched LLM backbones. With DeepSeek-v4-flash-0731, it achieves 93.69% symbolic accuracy on LSR-Transform, compared with 62.16% for SR-Scientist, and retains 72.97% accuracy on an anonymized variant that removes scientific descriptions and variable semantics, versus 39.64% for SR-Scientist. Under the same DeepSeek-v4-flash-0731 backbone, SRHarness also substantially outperforms Codex (72.97% vs. 20.72%) and reaches performance comparable to Codex with GPT-5.5, while simply providing Codex with the same scientific tools does not reproduce this advantage. These results show that effective agentic symbolic regression depends not only on models or tools, but also on structured runtime support for organizing scientific actions, accumulated hypotheses, and long-horizon search.

As the world moves towards sustainable energy sources, hydrogen (H2) can be treated as an eco-friendly alternative to fossil fuels due to its high energy density and zero carbon emissions. The water-gas shift (WGS) reaction is a widely used industrial process for hydrogen production by converting carbon monoxide and steam into hydrogen and carbon dioxide. However, occurrences like severe fouling, catalyst deterioration, and thermal runaway can hamper the reaction kinetics/process safety and decrease the yield of H2. These incidents are rare, and gathering process data under such abnormal conditions is challenging. In this work, we propose a physics-guided conditional diffusion model to generate realistic rare-event trajectories for the WGS reaction. The proposed model integrates a conditional denoising diffusion probabilistic model (CDDPM) with governing laws of the reaction to generate physically consistent process trajectories. The conditioning features allow the model to produce high-quality synthetic profiles for rare-event domains that are typically beyond the training regimes. The generated rare-event trajectories then augment the raw dataset for a balanced distribution between normal and abnormal conditions. We further propose a hazard score to assess the risk severity of the operating condition based on the operating trajectory. Deep learning models are trained with the augmented dataset to diagnose the health status of the reaction. Simulation results show that the proposed physics-guided diffusion model outperforms data-driven models in terms of the quality of synthetic data and diagnosis performance for rare events.

Many learning tasks map an input distribution to an output distribution. A natural way to model such an operator is to transform each input sample using a continuous function that may depend on the entire input distribution, and then take the distribution of the transformed samples. This defines a measure-dependent pushforward model and includes measure-theoretic formulations of transformers. We ask when such models can approximate arbitrary continuous operators between spaces of probability measures. We first show that universal approximation fails when atomic inputs are allowed: some continuous measure-to-measure operators that split or redistribute atomic mass cannot be approximated arbitrarily well by deterministic pushforward models. We then introduce the uniform level set condition, which requires a continuous measure-dependent scalarization whose shrinking level set neighborhoods carry uniformly vanishing mass over the input family. This condition is satisfied, in particular, by compact families of absolutely continuous measures. On every compact family satisfying this condition, we prove that any continuous measure-to-measure operator with outputs of finite $p$-th moment can be uniformly approximated, in the $p$-Wasserstein distance, by continuous measure-dependent pushforwards. Combining our theorem with existing approximation results for measure-dependent in-context maps yields universal approximation by measure-theoretic transformers. We also extend the framework to continuously-varying source measures, yielding a corresponding universality result for a class of pushforward models that are closely aligned with cross-attention architectures.

Dynamic routing creates severe load imbalance in large-scale expert-parallel Mixture-of-Experts (MoE) training, turning GPUs that host hot experts into stragglers. As each MoE layer waits for its slowest rank, these stragglers prolong the expert-parallel stage and reduce overall training efficiency. Existing expert-parallelism load-balancing (EPLB) systems commonly compute load-balancing plans on the CPU, incurring device–host data transfers and cross-rank synchronization that make scheduling at every layer and microbatch expensive. Their planning formulations also overlook the hierarchical communication costs of modern scale-up and scale-out GPU clusters. We present TopoEP, a GPU-native, topology-aware load-balancing system for large-scale MoE training. At each MoE layer and training microbatch, TopoEP converts the current routing result into hot-expert replication and token-rerouting decisions and executes the resulting plan without data-dependent host synchronization, reducing critical-path overhead. To generate these decisions, TopoEP uses a deterministic GPU solver that performs inter-node placement followed by intra-node refinement, allowing all ranks to independently produce bitwise-identical plans. On a 32-GPU NVIDIA H800 cluster, integrating TopoEP with Megatron-LM improves end-to-end training throughput by 6.2%–11.4% across three representative MoE models.

In self-supervised pretraining for Handwritten Text Recognition (HTR), pixel reconstruction methods outperform contrastive methods, unlike in natural-image classification. We argue that this difference follows from where discriminative signal lies in pixel space: for HTR, it is concentrated in high-variance directions and largely absent from low-variance ones. This predicts that objectives preserving high-variance pixel content will transfer best. We test six SSL methods from three families (pixel-grounded MIM, JEPA, and contrastive) under matched encoder, data, and evaluation protocols on six handwriting benchmarks across five languages. With full labels, pixel-groundrounded SSL achieves the lowest CER on every benchmark and both frozen probes, exposes per-position character information that other families recover only through the readout, and is the only family to benefit from pretraining on real handwriting. Pixel-grounded representations are also more label efficient. Across datasets, encoder alignment with the high-variance pixel subspace predicts CER within every method. With a pretrained LLM decoder, a frozen pixel-grounded encoder is competitive with fully fine-tuned supervised baselines; full fine-tuning achieves the lowest mean CER and ranks first or second on every benchmark. These results show that the value of pixel reconstruction depends on where discriminative signal lies in the input.

Autoregressive Vision-Language-Action (VLA) models offer a scalable path to robot learning, yet existing action tokenizers treat tokenization as a compression problem, producing representations that are semantically misaligned with the autoregressive backbone. We propose CATok, a causal action tokenizer that reframes tokenization as a causally structured generative process. CATok introduces a conditional annealing mechanism that extracts action tokens by progressively annealing a flow-matching process: each token is conditioned on all preceding tokens and encodes the residual reconstruction signal at a specific noise level, establishing a coarse-to-fine causal token space whose generative semantics are structurally aligned with autoregressive modeling. A token-conditioned flow-matching decoder built on Multimodal Diffusion Transformer (MMDiT) reconstructs continuous action chunks from these discrete tokens with the precision of hybrid diffusion-head architectures. This discrete bottleneck enforces knowledge insulation by design, cleanly separating high-level semantic reasoning from low-level motor execution without requiring explicit attention masking. Extensive evaluations across three simulation benchmarks and real-world robotic manipulation tasks demonstrate that CATok consistently surpasses existing tokenization methods in both reconstruction fidelity-compression tradeoff and inference efficiency, while improving VLA task success rate and training efficiency, establishing a high-performance, scalable foundation for purely autoregressive VLA systems.

Soft Actor-Critic (SAC) is widely used for entropy-regularised reinforcement learning with continuous action spaces, and practical implementations perform only a few actor steps towards an evolving target. In this work, we prove convergence guarantees when the target policy arises from policy mirror descent and compare it with the classical Gibbs target. We derive sufficient conditions for the strong convexity and smoothness of the actor objective, characterised by the curvature of the $Q$-function estimate through the Legendre differential operator, and establish an $\mathcal{O}\!\left(N^{-\frac{1}{5}}\right)$ best-iterate finite-time convergence rate up to actor and critic approximation errors. Moreover, the mirror-descent step size $\lambda$ directly controls the target drift and hence actor tracking error, whereas the analogous Gibbs bound contains a non-vanishing tracking term.

Leech-lattice quantization gives good quality at two bits per weight, but its codebooks hold more than 10^14 points, too many for a lookup table. Our earlier kernel expanded the codes at load time and read 4.804 bits per weight from GPU memory for 2 bits of code. We present Tetra, a new codebook on the same lattice. A 24-weight block still takes 48 bits, most of which index a 64-state trellis of the Golay code and one shared 16 KiB table. The kernel decodes a block with six table loads and two small lookups inside the matrix-vector product, and reads 2.148 bits per weight. For full models, we retrain one scale per matrix row, store the matrices that lose the most as 4-bit integers, and pay for them with 4-bit embedding tables. Our Qwen3-4B, 8B and 14B files hold 2.73, 2.70 and 2.73 bits per parameter over the whole model. They score 63.37, 69.58 and 75.66 on the full MMLU test set, 4.76, 4.21 and 2.46 points below 4-bit AWQ at 5.3 to 6.0 bits per parameter. They generate 113.8, 95.0 and 57.2 tokens per second in our engine. On GSM8K, through the served kernel, they lose 9.63, 4.62 and 3.26 points to FP16. At 4B our file scores 23.6 points above llama.cpp's IQ2_XXS (2.48 bits per parameter). Every number we measured for a table or figure comes from one NVIDIA L40S GPU. We preregistered the main experiments.

Patients often have several co-occurring clinical conditions, and the findings needed to identify and disambiguate them emerge over the course of a consultation. Evaluating clinical reasoning in this setting requires both multi-turn interaction and multi-label diagnosis. We introduce CLIMB, a benchmark in which a doctor model interviews a simulated patient to recover a ground truth set of co-occurring clinical conditions. Cases are synthesized from clinical decision algorithms and diagnostic datasets, grounding multimorbid presentations in structured clinical knowledge. Across six frontier and open models, none recovers the exact set of conditions in more than 10% of interactive cases. Diagnostic performance declines when conditions co-occur, even when models receive the full clinical record and the true number of conditions. Interaction reduces performance further. In controlled experiments, models behave like single-hypothesis trackers: they anchor on the diagnosis suggested by the opening findings, keep questioning around it, and recover a second condition mainly when a finding in view points to it. Questioning them further does not complete the set but adds mostly wrong diagnoses. We formalise this pattern with a theoretical reference model of single-hypothesis tracking. The benchmark, generator, and evaluation code are available at https://anonymous.4open.science/r/CLIMB-8340.

Flow matching (FM) learns generative dynamics through velocity regression. Geometric FM variants commonly assume a prior supported on the data manifold, requiring geometric knowledge that is often unavailable. Without such knowledge, low regression error alone does not guarantee manifold adherence. Adherence keeps generated samples within valid configurations and is empirically associated with better task performance. We introduce manifold-stable flow matching (MSFM), which can start from an arbitrary ambient prior, not necessarily supported on the manifold. Using tools from nonlinear dynamics, namely contraction theory, MSFM combines learned tangential transport with prescribed normal contraction. The construction uses analytical projectors for known manifolds and local affine proxies estimated by principal component analysis for unknown data geometry. By implementing contraction theory in both cases of known and unknown manifolds, we guarantee manifold invariance and transverse convergence to the manifold within a desired time window (e.g., one second). We derive a family of compatible probability paths and decompose the training loss into a learnable tangential term and a normal residual. An ellipse experiment attains a mean terminal off-manifold error of order $10^{-6}$. In Push-T robotic experiments, MSFM raises success from $74\%$ to $82\%$. In the Robomimic Square task, success increases from $60\%$ to $72\%$, while rotation-manifold deviation decreases from order $10^{-2}$ to $10^{-7}$. The MSFM terminal geometric errors are controlled by the chosen numerical tolerance. These results demonstrate stronger geometric adherence and higher observed task performance, supporting prescribed normal contraction as a complement to learned generative transport.

With limited annotation budgets, choosing which images to label determines how much a model improves. Data-selection methods that use features from a separately trained model, or scene descriptions written by vision-language models, have been successful, but those signals do not directly capture changes in the model being improved. The target model's own internal features reflect what it has learned so far and change with retraining, making them a natural cue for choosing the next training data. However, feature rarity alone does not reveal the errors that matter for performance. Here we link internal features to prediction errors and their expected impact on performance and select images for labeling and retraining without using labels for candidate images. We evaluated the method with an object detector on two datasets and two pairs of random seeds. Adding internal features improved the identification of prediction errors in 15 of 16 conditions. When performance was averaged over successive labeling rounds, the method outperformed selection based only on feature rarity in all four evaluation settings and ranked among the top two of six methods. With other conditions held fixed, performance after retraining was again higher than with rarity-based selection, even though the latter collected more errors. With longer retraining, the proposed method ranked first among six methods. These results suggest that linking a model's internal features to its errors and their effects on performance may help select training images that improve performance, thereby allowing the model's current state to guide which images are labeled next.

Effective central nervous system (CNS) intervention requires more than predicting biological responses: a scientific system must determine which intervention, observation, or additional evidence can improve a decision under uncertainty. However, in partially observed neuronal microenvironments, identical predicted outcomes may correspond to different exposure states, biological mechanisms, and intervention risks. We propose NeuronSifter, a decision-centric Agent-in-Twin framework, where an AI agent operates with an evolving biological digital twin rather than a fixed predictive model. NeuronSifter compiles interventions into state-conditional target-occupancy representations, propagates them through a mechanism-grounded World Action Model (WAM), and selects informative measurements based on their expected reduction in decision uncertainty. A shared probabilistic belief state couples intervention simulation, observation selection, and outcome assimilation, enabling the digital twin to continuously refine its decision-relevant representations through externally verified evidence. We evaluate NeuronSifter in Alzheimer's disease intervention scenarios using mechanistic neuronal microenvironment simulations and retrospective clinical endpoint references. The evaluation framework examines adaptive evidence acquisition, uncertainty-aware intervention planning, and robustness under model discrepancy through controlled environments and independent comparisons. NeuronSifter demonstrates how an AI system can actively identify missing evidence, refine biological hypotheses, and optimize intervention decisions under limited experimental budgets. Code is available at https://anonymous.4open.science/r/NeuronSifter.

State space models (SSMs) achieve efficient sequence processing because their affine state updates are closed under composition and can therefore be evaluated with an associative parallel scan. Nonlinear recurrent models can provide richer, state-dependent dynamics, but generally lose this compositional structure: parallel evaluation then requires iterative methods that repeatedly linearize and scan the recurrence. We ask, what state-dependent nonlinear dynamics can be designed to remain exactly composable? We answer by introducing RiccatiSSM, a nonlinear SSM, in which each state dimension follows an input-conditioned Riccati differential equation. Its quadratic state dependence makes the local Jacobian explicitly state-dependent, while its exact per-step flow under piecewise-constant inputs is a Möbius transformation. Since Möbius maps are closed under composition and compose through $2\times 2$ matrix multiplication, the complete nonlinear state trajectory can be evaluated exactly with a single associative parallel scan, without iterative linearization. We further derive a constrained parameterization that ensures bounded, contractive dynamics, and avoids poles in the fractional-linear state update. Across long-sequence classification, regression, and forecasting tasks, RiccatiSSM achieves competitive predictive performance while reducing runtime by $22{-}33\%$ compared to the nonlinear LrcSSM under matched architectures. These results demonstrate that state-dependent nonlinear dynamics can retain exact composability and be evaluated efficiently within a single parallel scan.

Large language models are trained with backpropagation, whose global gradient coordinates all layers but forces each to hold its activations and wait for the gradient to pass back through every deeper layer. Conventional local learning removes this update locking by training each module to predict the target through its own readout, but has not scaled to billion-parameter pretraining. We identify these private readouts as a key weakness, since they leave each module without information from deeper modules. We propose Shared-Output LOcal learning (SOLO), which replaces them with a shared, read-only copy of the final module's readout, the only one trained on the output of the whole network. Taken from the previous step, the copy transmits information from the final module without passing gradients between modules or reintroducing update locking. SOLO approaches backpropagation on Transformers of 340M to 2B parameters pretrained on 15B tokens, staying within one point in average zero-shot accuracy with a perplexity gap that narrows with scale. Readout ablations attribute SOLO's improvement over private readouts to sharing. Without update locking, each of p pipeline stages holds activations for O(1) micro-batches instead of O(p). The freed memory permits larger micro-batches, which reach up to 1.44x the best measured throughput of pipeline backpropagation on the same partition. To our knowledge, SOLO is the first local learning method to show such memory and throughput gains in billion-parameter language-model pretraining. Local learning thus becomes a practical alternative to backpropagation for large-scale pretraining.

Recently, deep neural networks on manifold-valued representations have garnered significant attention across various machine learning applications. One recent focus is the generalization of Euclidean fully connected (FC) and convolutional layers to non-Euclidean geometries. However, previous approaches typically focus on a few selected manifolds and rely on specific properties of the target manifold. In contrast, this work proposes a framework for constructing FC and convolutional layers over computationally tractable Riemannian spaces. This framework incorporates several previous FC layers across different geometries as special cases and is instantiated on ten representative manifolds, including three hyperbolic models, five geometries of the symmetric positive definite (SPD) manifold, and two Grassmannian perspectives. Experiments on different manifolds demonstrate the effectiveness and applicability of our approach. Code can be found at https://github.com/GitZH-Chen/RieTrans.

Reinforcement learning from verifiable rewards (RLVR) frequently reuses rollouts across multiple policy updates, increasing the mismatch between the current policy and the data-generating policy. We identify a sign-dependent gradient starvation problem in clipped policy optimization: clipping suppresses under-generated positive responses at the low-importance-weight tail while permitting severely over-generated negative responses to dominate the high-weight tail. To address this, we propose ReSPO (Reshaped Sequence Policy Optimization), which replaces clipping with a smooth, two-branch sequence-level kernel derived from an $\alpha$-divergence variational objective and an exponential variance-control tilt. The positive branch preserves a nonzero gradient weight for under-generated positive responses, while the negative branch suppresses heavily over-generated negative responses. We demonstrate that ReSPO effectively learns from long positive reasoning trajectories during early training, even when accumulated policy drift relegates them to the low-importance-weight tail. On dense and MoE Qwen3 models, ReSPO accelerates early optimization, improves final training scores, and achieves higher held-out benchmark performance under a rollout reuse, validating our approach on importance-weight tail control in off-policy learning.

Estimating conditional statistics and learning representations of a population of conditional distributions are central problems in many data-driven applications, including uncertainty quantification and dynamical systems analysis. Conditional mean operators (CMOs), a class of linear operators between function spaces, resolve these objectives by providing access to a broad class of conditional statistics. However, existing methods typically estimate each CMO independently or constrain it to prespecified function spaces, thereby preventing the exploitation of shared structure across related distributions. In this work, we posit that related CMOs share finite-dimensional input and output function spaces, and are specialized for each task with a linear operator mapping these spaces. Based on this hypothesis, we introduce MTL-CMO, a multi-task framework that jointly learns shared function spaces and task-specific operators across multiple datasets. We further introduce T-CMO, a transfer learning method that reuses the shared spaces to estimate, in closed form, the operator of a new conditional distribution. We establish statistical guarantees quantifying the benefits of jointly learning the shared function spaces. Our experiments demonstrate that learning shared function spaces improves uncertainty quantification across a broad range of conditional distributions and, when applied to Langevin and plasma dynamics, yields compact representations of complex dynamics that retain physically meaningful information and enable parameter identification.

Modern LLMs are increasingly capable as autonomous agents, but they follow sequential interaction cycles: read, think, reply or call tools, repeat. Many real-world use cases are not sequential: voice assistants, embodied agents, and monitoring systems receive new inputs while they think or perform another task. Modern LLMs address this with specialized architectures for voice interaction and video streams, VLAs for robot control, asynchronous tool calling for API usage, and others. In this work, we generalize from different asynchronous tasks to general asynchronous agents that can adapt to different types of concurrency. To achieve this, we develop an asynchronous LLM framework that lets users (or the agents themselves) define inference coroutines with overlapping memory states. We showcase that Qwen 3.x models are capable of asynchronous operation for streaming video understanding, videogames, and monitoring, without task-specific training.

Reinforcement Learning-based post-training of Large Language Models (LLM) has been successfully applied to improve their reasoning capabilities. Existing pipelines primarily finetune LLMs on a fixed pool of problems specified prior to training using the GRPO loss. This is fundamentally limiting, as learning signal arises only when policy rollouts mix successes and failures, causing the useful portion of any fixed pool to quickly become stale as the model improves. To address this, we propose frontier learning, an open-ended post-training approach in which procedural generators are used online to continually produce informative training problems. It treats the generator's task-specific parameters as a search space and uses a regret signal to prioritize and explore frontier difficulty levels in order to focus training at the edge of the model's evolving reasoning capabilities. Across several reasoning tasks and model families, our approach consistently achieves higher relative gains over fixed-pool baselines, demonstrating that effective post-training requires not only selecting useful problems, but continually generating them at the edge of capability.

This paper studies convex optimization when the gradient cannot be evaluated exactly, but only approximated by a hierarchy of algorithms whose compute grows like $\delta^{-\gamma}$ in the accuracy $\delta$. When $\gamma>2$, falling into the Harder-Than-Monte-Carlo (HTMC) regime, the price of accuracy outruns the variance reduction that Monte Carlo would buy and we show that minimizing a loss function costs no more, up to a factor depending only on $\gamma$, than a single evaluation of its gradient at the accuracy the problem demands. A randomized multilevel oracle replaces the deterministic approximation of accuracy $\delta$ by an unbiased estimator of it, whose variance $\sigma^2$ becomes a second, independently priced dial: the cost of one call drops from $\delta^{-\gamma}$ to $\delta^{2-\gamma}\sigma^{-2}$. Plain inexact gradient descent driven by that oracle reaches loss $\varepsilon$ at expected compute $\Theta(\varepsilon^{-\gamma})$ in the convex case, against $\Theta(\varepsilon^{-(\gamma+1)})$ for the same method run at a fixed accuracy: randomization buys a full power of $\varepsilon$. Under $\mu$-strong convexity the exponent halves, to $\varepsilon^{-\gamma/2}$, because the iterates settle at a noise floor and the bias budget relaxes accordingly. Both bounds are independent of the step size, and hence of the smoothness constant, and we show that the cost is a functional of the underlying gradient flow rather than of any discretization of it.

When pricing agents meet repeatedly on a platform, the platform decides who faces whom. We ask whether that choice moves the prices the agents learn, and whether a rise comes with learned punishment. In a pre-registered randomised experiment in the Bertrand duopoly of Calvano et al., each agent's price is set by a tabular Q-learning module, not by the small language model attached to it, and we randomise whether each agent keeps its partner, sees its rival's prices and can send messages. Keeping the same partner raises the level of profits, averaged over training, by 0.27 of the gap between competitive and monopoly profit (95% CI 0.20 to 0.35, all twenty paired runs positive), our registered primary result, and the resting price by 0.17 of the Nash-to-monopoly range (post hoc). A plain tabular learner reproduces the effect in all 25 further blocks, and there one permanent partner raises the level more than about three do (+0.23 against +0.05, exploratory). Where rival prices are hidden, the price-setting module cannot see a cut, so cannot punish it, yet the resting price rises as much and the rise lasts to the end of training, while with visible rivals it shrinks with longer training (post hoc). Where the rival is visible, a static best responder accounts for a third to a half of what a forced-deviation probe reads as punishment, on the starts where the rival can see the cut, and net of it the registered test of learned punishment is inconclusive. A test that looks only for punishment would thus miss the rise where the rival is hidden, while a check for profitable deviations flags most of those prices (post hoc). In an exploratory extension, untrained Qwen2.5 7B and 14B models under one prompt show the effect when the rival's price is left out of the prompt and inconsistently when it is shown, the 7B result replicating on fresh blocks, while two other model families show none.