NeurIPS 2026 Papers

Modern automatic speech recognition (ASR) systems have achieved remarkable accuracy by scaling model depth and capacity, but at the cost of substantial memory and computation. On edge devices where ASR is often most needed, such as watches and glasses, deploying such large models is infeasible due to extremely constrained resources. This raises a fundamental question: can we achieve high representational power in deep ASR models without scaling up parameterization? In this work, we revisit the role of depth and identify layer-wise representational dynamics, in which most layers learn functionally similar transformations. Motivated by this insight, we propose a block-recurrent ASR architecture that replaces parameterized depth with a small set of recurrent blocks, each consisting of weight-shared layers, thereby preserving effective depth while drastically reducing model size. Through representation-guided grouping and knowledge distillation, block-recurrent models retain the accuracy of large reference models while using an order of magnitude fewer parameters. Across Open ASR Leaderboard benchmarks, a model with only two shared blocks recovers $97.4$% of the accuracy of large models on average. The compact backbone further enables efficient on-device personalization through lightweight MoE-LoRA adaptation, allowing user-specific models to match or surpass the accuracy of general-purpose ASR systems with significantly lower inference cost. Together, these results show that high-quality ASR can be achieved without large parameterization, and that structured recurrence provides a principled path towards ultra-compressed, high-efficiency, and personalized speech recognition on the edge.

Language model loss follows remarkably regular scaling laws over model and data size, yet it remains unclear why the aggregate loss should exhibit a power-law form. Existing explanations often attribute this regularity to a heavy-tailed spectrum of pattern difficulty in natural language, but this view has not been directly validated at token-level granularity in large-scale real-data training. We present a token-level framework that deconstructs scaling laws into localized learning events of individual contextualized tokens. By fitting token loss trajectories with sigmoids, we show that token learning is concentrated in localized transitions, giving rise to a learning-time spectrum that dominates the scaling-law shape. Across more than one hundred pre-training runs on large and diverse real-language corpora with modern LLM architectures, scaling up to 6B parameters and 300B training tokens, the measured learning-time spectrum quantitatively reconstructs the validation loss derivative along the training-step $T$, data-scale $D$, and model-scale $M$ axes. We further show that the same signal is actionable: by reshaping the training distribution according to when tokens become learnable, we alter the optimization trajectory and achieve 11\% faster validation-loss reduction. These results provide direct empirical evidence that scaling laws are governed primarily by the distribution of token-level learning times, and that this distribution can be used not only to explain scaling behavior but also to improve training performance.

The task of \textit{dataset distillation} aims to find a small set of synthetic images such that training a model on them reproduces the performance of the same model trained on a much larger dataset of real samples. Existing distillation methods focus on synthesizing datasets that enable training \textit{randomly initialized} models. In contrast, state-of-the-art vision approaches are increasingly building on large, pre-trained self-supervised models rather than training from scratch. In this paper, we investigate the problem of distilling datasets that enable us to optimally train \textit{linear probes} on top of such large, pre-trained vision models. We introduce a method of dataset distillation for this task called \textit{Linear Gradient Matching} that optimizes the synthetic images such that, when passed through a pre-trained feature extractor, they induce gradients in the linear classifier similar to those produced by the real data. Our method yields synthetic data that outperform all real-image baselines and, remarkably, generalize across pre-trained vision models, enabling us, for instance, to train a linear CLIP probe that performs competitively using a dataset distilled via a DINO backbone. Further, we show that our distilled datasets are exceptionally effective for fine-grained classification and provide a valuable tool for model interpretability, predicting, among other things, how similar two models' embedding spaces are under the platonic representation hypothesis or whether a model is sensitive to spurious correlations in adversarial datasets.

Cold-start exemplar-free class-incremental learning requires learning a growing set of classes without replay, external pretraining, or a large initial task. Existing cold-start methods typically either train the backbone throughout the stream and compensate for semantic drift, or freeze a backbone after the first task, producing features biased toward the initial classes. These choices also create a computational tension: drift-compensation methods require repeated backbone training and increasingly expensive updates as the task horizon grows, while frozen-backbone methods are cheap but weak under cold start. We study a third option: a feature extractor that is never fit to image data at all. We propose CIRCLE, a class-incremental classifier built from fixed bidirectional two-dimensional reservoir features, adapted from BiRC2D for image classification, and streaming linear discriminant analysis heads. CIRCLE groups multiple random reservoir instantiations into feature ensembles and averages the softmax outputs of independent SLDA heads, yielding a tunable bias-variance tradeoff between richer random features and prediction-level ensembling. Because the feature extractor is fixed and the head admits streaming closed-form updates, CIRCLE performs sample-wise training without replay, task-boundary information, or backbone backpropagation. On CIFAR-100, TinyImageNet, ImageNet-Subset, and ImageNet-1k, CIRCLE is competitive at 10-20 task splits and substantially outperforms strong CS-EFCIL baselines at 50, 100, and 500 task splits, while training much faster than trained-backbone drift-compensation methods. Ablations show that the BiRC2D-style extractor, SLDA head, and balanced feature/prediction ensembling each contribute to the final performance.

Classical scaling laws for language model pretraining balance model size against training dataset size under a fixed compute budget, assuming abundant data and a single pass over the corpus. As training compute grows faster than the supply of natural language data, pretraining is likely to enter a data-constrained, compute-rich regime where models train for multiple epochs over a finite dataset. We study data-constrained pretraining along two axes, regularization and scaling. For regularization, we study masked-input regularization (MIR), an auxiliary next-token prediction loss on randomly masked inputs. MIR tests whether the random masking central to diffusion language models can benefit autoregressive pretraining without architectural changes or inference overhead. Across 72M to 1.4B parameter models, we find that MIR added on top of strong weight decay improves validation loss over autoregressive strong-weight-decay-only models, with downstream gains at 1.4B. For scaling, we propose SoftQ, a scaling law that couples model size and data size to capture their interaction under repeated data. Classical alternatives such as the Chinchilla law use an additive form that decouples these terms, making them misspecified in the data-constrained regime. We find that SoftQ fits data-constrained experiments substantially better than these alternatives, and estimates MIR's gains as equivalent to roughly 1.3 times as much unique training data.

Constrained decoding is essential for serving LLMs, ensuring that generated outputs follow specific structures such as JSON schema-formatted function calls. Existing systems are designed for autoregressive models and assume left-to-right generation, masking out invalid next tokens at each step. Diffusion language models, however, break this assumption: they sample multiple positions simultaneously from a fully-factorized mean-field distribution at each denoising step. In this paper, we present an exact and tractable algorithm for sampling from the constrained mean-field posterior under any constraint expressible as a finite automaton. Viewing finite automata as graphical models, we obtain tractable representations of the constrained distribution that enable efficient inference. The approach guarantees constraint satisfaction by construction, supports both greedy and sampling-based decoding, and is compatible with parallel and block-wise decoding under arbitrary remasking schedules. Applying depth-reduction techniques from arithmetic circuit theory, we further reduce sampling depth from linear to logarithmic in the sequence length. Empirical evaluations on Dream-7B and LLaDA-8B show substantial accuracy gains across various tasks including function calling (xLAM, BFCL), planning (Sudoku, Countdown), text-to-SQL (Spider), and math reasoning (GSM-Symbolic), with little inference overhead relative to unconstrained decoding. For example, on BFCL-Live, our approach improves Dream-7B's greedy decoding accuracy from 63.9% to 71.5%, and stochastic sampling accuracy from 22.3% to 69.0%, where the unconstrained baseline collapses, with under 5% wall-clock overhead.

Diffusion language models (DLMs) have emerged as a promising alternative to autoregressive (AR) models for language modeling, allowing flexible generation order and parallel generation of multiple tokens. However, this flexibility introduces a challenge absent in AR models: the \emph{decoding strategy}---which determines the order and number of tokens generated at each iteration---critically affects sampling efficiency. Among decoding strategies explored in practice, confidence-based methods, which adaptively select which and how many tokens to unmask based on prediction confidence, have shown strong empirical performance. Despite this success, our theoretical understanding of confidence-based decoding remains limited. In this work, we develop the first theoretical analysis framework for confidence-based decoding in DLMs. We focus on an entropy sum-based strategy that continues unmasking tokens within each iteration until the cumulative entropy exceeds a threshold, and show that it achieves $\varepsilon$-accurate sampling in KL divergence with an expected number of iterations $\widetilde O(H(X_0)/\varepsilon)$, where $H(X_0)$ denotes the entropy of the target data distribution. Notably, this strategy yields substantial sampling acceleration when the data distribution has low entropy relative to the sequence length, while automatically adapting to the intrinsic complexity of data without requiring prior knowledge or hyperparameter tuning. Overall, our results provide a theoretical foundation for confidence-based decoding and may inform the design of more efficient decoding strategies for DLMs.

Chunked prefill has become a widely adopted serving strategy for long-context large language models, but efficient attention computation in this regime remains challenging. Existing sparse attention methods are primarily designed for one-shot prefill and do not translate efficiently to chunked prefill: block-sparse kernels lose efficiency when the query length is limited by the chunk size, while fine-grained pattern search becomes costly when repeated over the accumulated KV cache at every chunk. QUOKA, a recent method that directly targets chunked prefill, avoids sparse-kernel overhead but relies on query-subsampled, token-level KV selection, which can miss query-specific KV entries and introduce explicit KV-copy overhead. To address these limitations, we propose CompactAttention, a chunked-prefill attention mechanism based on Block-Union KV Selection. CompactAttention treats 2D block-sparse masks as KV-selection signals rather than direct sparse-kernel execution plans, and converts them into GQA-aware per-group KV block tables through Q-block union and intra-group union. This construction produces the minimal block tables that preserve all KV blocks selected by the input masks under paged execution constraints, enabling selected KV blocks to be accessed in place without explicit KV compaction. On LLaMA-3.1-8B-Instruct, CompactAttention maintains accuracy close to dense attention on the RULER benchmark while delivering up to 2.72$\times$ attention speedup at 128K context length under chunked prefill.

Scaling Transformers to long contexts is constrained by the quadratic cost of self-attention and the linear growth of key-value cache memory transfer. Sparse attention mitigates this by retrieving only relevant tokens, but current approaches either require large-scale training or, within the training-free regime, rely on semantically coarse heuristics or expensive clustering that is difficult to update efficiently during decoding. We introduce CommunityKV, a framework that formulates sparse attention as a community detection problem. CommunityKV constructs a token graph from the $QK^T$ scores already computed during standard prefill, and partitions the graph into communities to enable retrieval of semantically coherent token groups. A local update rule assigns newly generated tokens to communities in constant time, enabling sparse retrieval throughout streaming decoding without global re-partitioning. We evaluate CommunityKV on open weight models and long context benchmarks, demonstrating that our approach achieves up to $1.86\times$ higher decoding throughput than exact FlashAttention-2 with negligible accuracy degradation.

Parameter-efficient fine-tuning methods, such as LoRA, offer a practical way to adapt large vision and language models to client tasks. However, this becomes particularly challenging under task-level heterogeneity in federated deployments. In this regime, personalization requires balancing general knowledge with personalized knowledge, yet existing approaches largely rely on heuristic mixing rules and lack theoretical justification. Moreover, prior model merging approaches are also computation and communication intensive, making the process inefficient in federated settings. In this work, we propose Potara, a principled framework for federated personalization that constructs a personalized model for each client by merging two complementary models: (i) a federated model capturing general knowledge, and (ii) a local model capturing personalized knowledge. Through the construct of linear mode connectivity, we show that the expected task loss admits a variance trace upper bound, whose minimization yields closed-form optimal mixing weights that guarantee a tighter bound for the merged model than for either the federated or local model alone. Experiments on vision and language benchmarks show that Potara consistently improves personalization while reducing communication, leading to a strong performance-communication trade-off.

Decentralized training enables large-model training over low-end GPUs and internet-grade connections, but communication along both data-parallel and pipeline-parallel axes becomes the primary bottleneck. We study post-pretraining adaptation in this setting. We propose an asynchronous two-circuit system: a fast compressed training circuit drives throughput using activation masking for pipeline-parallel (PP) transfer and compressed data-parallel (DP) synchronization, while a slow anchor circuit runs occasional unmasked forward--backward passes off the critical path. Then, we introduce a spectral correction optimizer that uses these delayed anchor priors to denoise masked gradients without blocking the fast stream. Although prior work has found aggressive activation compression unreliable, we show that masking supports post-pretraining adaptation at high compression rates when anchored this way. Pipeline-parallel compression alone yields up to a 9x throughput gain, and combining it with data-parallel compression increases beyond 40x over internet-grade 200Mbps connections, while matching dense uncompressed performance across domain adaptation and continual pretraining.

Video Language Models enable AI systems to understand temporal dynamics in videos. To fit within the maximum context window constraint, current methods use keyframe sampling which often misses both macro-level events and micro-level details due to the sparse temporal coverage. Furthermore, processing full images and their tokens for each frame incurs substantial computational overhead. We address these limitations by leveraging video codec primitives (specifically motion vectors and residuals) which natively encode video redundancy and sparsity without requiring expensive full-image encoding for most frames. To this end, we introduce lightweight transformer-based encoders that aggregate codec primitives and align their representations with image encoder embeddings through a pre-training strategy that accelerates convergence during end-to-end fine-tuning. Our approach, *CoPE-VideoLM*, reduces the time-to-first-token by up to 86% and token usage by up to 93% compared to standard VideoLMs. Moreover, by varying the keyframe and codec primitive densities, we maintain or exceed performance compared on 14 diverse video understanding benchmarks, spanning general question answering, temporal and motion reasoning, long-form understanding, and spatial scene understanding.

Prior-fitted networks (PFNs) are a promising class of tabular foundation models that perform in-context learning, whereby the entire labelled training set is supplied as context, and predictions for test queries are produced in a single forward pass. However, the quadratically scaling self-attention mechanism in many PFN architectures makes inference prohibitive for very large training datasets. We propose CRUMB (Clustered Retrieval Using Minimised-MMD Batching), a three-stage inference wrapper that (i) clusters the test queries, (ii) selects a small, distributionally matched training subset for each cluster by greedily minimising the maximum mean discrepancy (MMD), and (iii) runs exact PFN inference on each reduced-context batch. CRUMB is architecture-agnostic and requires no retraining. On the 51-dataset TabArena benchmark, evaluated across three PFN architectures (TabPFNv2, TabICLv1, TabICLv2), we show that CRUMB outperforms similar state-of-the-art context selection strategies. We also show that CRUMB is resilient to covariate drift, as the MMD-minimisation step naturally helps align the training context distribution to match the current test batch distributions.

Chiplet-based GPUs are a promising alternative to monolithic designs for scaling compute and memory resources in machine learning (ML) workloads. However, distributing computation and memory across multiple chiplets introduces a hierarchical Non-Uniform Memory Access (NUMA) environment, where performance depends on the alignment between kernel execution and data placement. In deep neural network (DNN) inference, kernels execute repeatedly over shared weights and activations, and poor alignment leads to sustained inter-chiplet traffic that degrades performance. Existing approaches optimize scheduling or data placement in isolation, failing to coordinate these decisions across dependent kernel executions. We present CHoRD (Chiplet-Hierarchy-oriented Residency and Dispatch), a compiler-assisted, profile-guided framework for joint scheduling and data placement in chiplet-based GPUs. CHoRD reconstructs program structure from compiled kernels and host code, and supports two modes: a static mode that derives scheduling and placement decisions without profiling, and a profiling mode that performs a single lightweight run to collect execution statistics. Using this information, CHoRD determines chiplet-level kernel mapping, enables safe inter-kernel overlap, improves intra-chiplet parallelism, and assigns data to chiplets based on access patterns. These coordinated decisions reduce inter-chiplet communication and improve compute utilization. We evaluate CHoRD on 4-chiplet and 8-chiplet MGPUSim substrates across nine CNN and transformer-based inference workloads. On the 4-chiplet system, CHoRD-static and CHoRD-profiled achieve 2.57× and 2.94× geometric-mean end-to-end speedups over RR, and 1.46× and 1.66× over the placement-only CLAP baseline, respectively, with benefits increasing as chiplet count scales to 8.

Can we trust evaluation scores to capture an LLM's true real-world performance? *Certifiable* evaluation answers this question by providing guarantee for LLM evaluation. In particular, existing methods sequentially curate evaluation samples and keep updating confidence intervals (CIs) that cover the true performance with high probability (e.g., 95%) until some conditions are satisfied, e.g., the CI width reaches a target precision. However, existing methods are not generally anytime-valid: the claimed coverage (e.g., 95%) may fail when CIs are repeatedly updated and used to decide when to stop, leaving a *gap* between theoretical rigor and practice. This paper bridges this gap by proposing **Celeus**, a **C**ertifiable framework for **E**fficient **L**LM evaluation, which leverages ***E***-processes to build anytime-valid CIs. Concretely, we propose signals that combine two ingredients: (i) **U**ncertainty-guided sampling to select informative samples for evaluation, and (ii) **S**urrogate-assisted approximations for unevaluated samples. We prove that such signals remain unbiased for the evaluation score conditional on the past, enabling statistically grounded and anytime-valid $e$-process CIs. More importantly, the two ingredients reduce estimation variance and help reach the target precision with fewer evaluated samples. We also prove that CIs obtained by **Celeus** can shrink at a near-parametric rate up to logarithmic factors and analyze the oracle variance-optimal sampling rule that motivates the empirical uncertainty-guided one. Experiments show that **Celeus** reaches the target precision using 54-62% fewer evaluated samples than baselines, while preserving anytime-valid coverage.

Token-level distillation assumes a shared categorical support, an assumption violated when teacher and student tokenizers segment the same text differently. Existing cross-tokenizer methods sidestep this mismatch by aligning token spaces, matching likelihoods along tokenizer-specific paths, or filtering shared spans, all of which can approximate the original distributional objective or discard probability mass. We argue that decoded UTF-8 bytes are common observable events shared by every text tokenizer, while token boundaries are tokenizer-specific latent transitions. Building on this view, we introduce \emph{ByteDistill}, which compares the conditional distribution of the next observable byte rather than aligning token identities. Its loss, \emph{Byte-Level Distribution Alignment} (BLDA), projects each native softmax into a 256-way observable next-byte distribution inside aligned decoded chunks, marginalizing token-ending mass as a latent boundary transition along the observed tokenization path. ByteDistill consistently outperforms supervised fine-tuning and prior cross-tokenizer baselines in both off-policy and on-policy settings, with off-policy training improving GSM8K accuracy by $1.75$ percentage points over the current state of the art.

Hierarchical Gaussian Process (H-GP) models divide problems into different subtasks, allowing for different components to address each part, making them well-suited for problems with inherent compositional structure. However, existing H-GP frameworks typically employ one-way information sharing — either top-down or bottom-up — which limits sample efficiency and slows convergence. We propose Bidirectional Information Flow (BIF), which establishes continuous two-way communication. BIF retains the modular structure of hierarchical models — the parent conditions its own posterior on child summaries, treating them as structured priors — while introducing top-down feedback to softly decompose environment observations from the parent into sub-responses. This mutual exchange improves sample efficiency, enables robust training, and allows modular reuse of learned subtask models. We prove analytically the regret of a GP with a learned kernel scales linearly with the mismatch to the true kernel, tightening in the hierarchical case to the sum of child-level errors. Ablation shows removing the downward pathway collapses child $R^2$ by up to 58\%. Across synthetic, neurostimulation, and HPO benchmarks, BIF achieves up to 4× higher parent $R^2$ and $\sim$100\% AUC improvement over vanilla GPBO, and outscores all hierarchical state-of-the-art methods on child $R^2$ given the correct acquisition function, while supporting modular child transfer to novel composite tasks.

Recent advances in gigapixel-level imaging have brought High-Resolution Wide shots to the forefront of research. However, these images present significant challenges: extreme sparsity of foreground, gigapixel-level resolutions and interleaving of foreground and background. This causes traditional detectors and attention mechanisms to be hindered by background, resulting in inefficiency and inaccuracy. To tackle this problem, we propose BiHashFormer, a dual sparsified branch transformer built on hashing. RoI Selector will predict the foreground proportion for top-k selecting target-containing windows. The compression branch processes windows selected and refines attention into HashAttention by discarding half of the key–value pairs, reducing cost and improving efficiency. The compensation branch uses HashMiner to perform low-cost hash searches in the remaining regions, recovering features of missed objects. Experiments confirm that the selection relationship in HashAttention is one-way and reveal the preference of different queries for keys. In experiments on the gigapixel benchmark PANDA, BiHashFormer reduces 37.5\% backbone FLOPs while improving $\text{AP}_{50}$ to 81.0\%. Moreover, this parallel dual-branch structure allows the model to avoid inference delays caused by inter-branch dependencies, while achieving much greater stability than single-branch models, specially with low rates of RoI selection.

Online bipartite matching, where agents are known in advance but items arrive sequentially and must be irrevocably assigned, is fundamental to problems ranging from ride-sharing to online advertising. When agents belong to classes such as demographic groups or geographic regions, fairness demands equitable treatment across these groups. Recent work introduced class envy-freeness (CEF), a natural extension of the classical fair division notion: an algorithm is $\alpha$-CEF if each class receives value at least an $\alpha$ fraction of what it could extract from any other class's bundle. However, all known algorithms achieving constant-factor CEF guarantees attain utilitarian social welfare (total matching value) of at most $\frac{1}{2}$ times the optimum, far below the $1-\frac{1}{e} \approx 0.632$ achievable without fairness constraints. We resolve the open question of whether fairness necessitates this efficiency loss, by introducing threshold-based algorithms parameterized by $\gamma \in [0,1]$ that equalize allocations across classes until threshold $\gamma$, then maximize efficiency. For divisible matching, this yields simultaneous $(1-e^{-\gamma})$-CEF and $(1 - \frac{e^{\gamma-1}}{\gamma+1})$-USW guarantees; for indivisible matching, $\frac{\gamma}{2}$-CEF with the same USW. Setting $\gamma > 0$ produces the first algorithms beating $\frac{1}{2}$-USW while maintaining constant CEF. We complement this with a novel upper bound construction, proving no non-wasteful $\alpha$-CEF algorithm can exceed $\frac{1 +\alpha - e^{\alpha-1}}{1+\alpha}$-USW and correcting prior bounds that were vacuous for $\alpha < 0.58$. Our upper bound nearly matches our algorithms' performance, giving the first substantive characterization of the price of fairness in online class matching.