Papers

Binary diffusion models typically require a large number of function evaluations (NFEs) to generate high-quality samples, making practical inference computationally expensive. Reducing NFEs while preserving sample quality without distillation or additional training remains a significant challenge. Existing binary diffusion models define a discrete one-step forward path and then derive the reverse posterior. In low-NFE settings requiring cross-step sampling, they approximate the true multi-step likelihood with a single-step likelihood transition, which severely degrades sample quality. To address this fundamental limitation and decouple the generative dynamics from fixed discrete time steps, we propose Bernoulli Flow Models (BFM). Rather than relying on sequential one-step Markov diffusion chains, BFM defines a unified continuous global Bernoulli probability flow path between data distributions and pure noise, from which we derive analytical closed-form posterior transitions over arbitrary time intervals. Consequently, reducing the inference NFE is no longer an approximation based on skipping discrete steps; it only requires re-evaluating the analytical posterior over a new time grid. This eliminates the structural training-inference mismatch inherent to discrete chains and yields self-consistent low-NFE sampling. Experiments show that BFM is highly robust to aggressive NFE reduction. On LSUN Churches 256x256, a BFM trained with 256 steps achieves an FID of 9.22 using only 16 sampling steps, whereas the state-of-the-art discrete baseline degrades to 204.10. BFM also remains competitive with continuous and discrete generative baselines under standard full-step inference. These results establish BFM as a theoretically rigorous, self-consistent, and practically effective framework for fast binary data generation.

Cross-model KV-cache reuse remains a key challenge in modern LLM serving. Coding agents and multi-model systems increasingly route a shared context across models: a user may switch models mid-session, or a cascade may escalate a difficult query. Because KV caches contain model-specific representations, each switch typically forces the receiving model to prefill the entire context from scratch. Recent work shows that closed-form linear maps can translate KV caches between models in the same family, but transfer accuracy degrades as the model-size gap widens. In this paper, we establish that these transfer failures are concentrated in a small subset of information-dense tokens. To bridge this gap, we introduce RaReCache, a framework that enables a large target model to decode accurately from a cache prefilled by a much smaller source via selective recomputation. RaReCache identifies these critical positions using a novel rank disagreement metric, scoring each token by the energy of its mapped KV in output directions weakly supported by the calibration data. Across two model families and five benchmarks, on a 23x parameter gap (Qwen3-0.6B to 14B) recomputing just 30% of positions retains 95-99% of the target accuracy, whereas on a 8.8x gap (Llama3-8B to 70B), recomputing 40% retains 96.5% of the target accuracy. RaReCache largely removes sensitivity to source-model size, and achieves up to a 3.04x prefill speedup. For online serving, it handles 1.8x the request throughput of target prefill on a single GPU, and at the target's saturation load, reduces median and 99th-percentile time-to-first-token (TTFT) by 5.0x and 6.4x respectively, with a 30% recompute budget. RaReCache establishes an efficient serving paradigm where small models prefill on behalf of massive targets, enabling large models to recompute only critical tokens, drastically reducing prefill latency.

Language models (LMs) are commonly trained with Reinforcement Learning with Verifiable Rewards (RLVR) to enhance their reasoning capabilities. However, since RLVR does not explicitly account for calibration during training, it can lead to severe calibration degradation, including overconfidence. Recent calibration-aware training methods for LMs, which incorporate objectives for uncertainty estimation into training, improve calibration but still exhibit overconfidence under distribution shift, while sacrificing reasoning performance. To this end, we propose RL-ARC, a calibration-aware training framework that jointly leverages reasoning confidence and answer confidence. Specifically, RL-ARC leverages reasoning confidence as an auxiliary signal for calibrating answer confidence, applying it as reasoning-guided regularization for correct cases and as an overconfidence penalty for incorrect cases. Comprehensive results across ID and OOD settings show that, beyond improving calibration, RL-ARC enables reasoning models to adaptively estimate confidence based on the given question without substantially sacrificing reasoning performance, thereby highlighting the importance of reasoning confidence for training reliable reasoning models.

Generative conformal prediction builds uncertainty sets from samples of a conditional generator, which are efficient only when the samples represent the response distribution well. This can require many samples, each of which can be costly, as in large diffusion models and scientific simulators, so the sampling budget must be used efficiently. Existing methods draw the same number of samples at every input, wasting samples where the response distribution is simple and undersampling where it is complex, which inflates sets and leaves those inputs under-covered. We propose CASA (Conformal Adaptive Sample Allocation), which characterizes the marginal value of an additional sample and allocates samples across inputs to minimize the expected set size subject to marginal coverage and an average sampling budget. Theoretical analysis shows that adaptive allocation yields smaller sets than a fixed count at the same budget: a missed mode forces a radius that spans the gap between modes, and even oracle radius cannot compensate for it. On synthetic and real tasks, CASA produces substantially smaller sets at the same budget, often improves conditional coverage, and complements existing radius-adaptive methods.

Mixture-of-Experts (MoE) architectures have become essential for scaling large language models, with recent work demonstrating the benefits of fine-grained expert designs. Training such models from scratch is expensive, and sparse upcycling from pre-trained dense models is an attractive alternative. However, we identify a structural pathology of fine-grained upcycling: when fine-grained experts are derived from a single source model, naive routing collapses and downstream accuracy drops to near-random (e.g., on Qwen3-1.7B, Drop-Upcycling-fine-grained reaches only 23.2% average accuracy across 15 benchmarks, essentially matching from-scratch training at 22.2%, while the same method's coarse-grained variant reaches 50.2%). We propose DivMoE, the first framework achieving fine-grained MoE Upcycling with structurally-balanced routing. DivMoE introduces domain-specialized fine-grained expert initialization, deriving experts from dense models that have undergone domain-adaptive continual pre-training, and diversity-constrained routing, a hard structural constraint guaranteeing that each token activates experts from distinct domain groups. Across two base models and 15 benchmarks, DivMoE consistently outperforms six upcycling baselines (55.6% vs. 51.6% for the strongest baseline on Qwen3-1.7B) and strictly improves over the dense base model on every benchmark after Stage 2 continual pre-training – closing the regression gap that has plagued prior fine-grained upcycling. After supervised fine-tuning on a public reasoning mixture, our 12B-parameter DivMoE model matches Moonlight-MoE (16B) at 64.5% average accuracy while outperforming a controlled NVIDIA-Upcycling baseline by 6.1 percentage points.

In diffusion transformers, low-rank branches can mitigate 4-bit weight–activation (W4A4) post-training quantization (PTQ) loss by decomposing each weight into a low-bit residual and a high-precision low-rank component. Existing low-rank PTQ approaches, however, either optimize low-rank compensation and residual quantization separately, often requiring higher ranks, or rely on second-order weight updates without explicitly modeling activation quantization error, which becomes particularly pronounced under 4-bit quantization. To address these limitations, we present HSVDQuant, a unified framework modeling low-rank-assisted W4A4 PTQ as a coupled calibration problem and deriving optimization-based solvers from the joint objective. Eliminating the output-side low-rank factor yields a deflated Hessian that discounts residual errors already captured by the low-rank component, while an activation-noise surrogate is incorporated to suppress activation quantization error. Across five diffusion backbones, rank-4 HSVDQuant consistently outperforms rank-4 SVDQuant in PSNR and LPIPS. It further surpasses rank-32 SVDQuant on SANA-1.6B, FLUX.1-schnell, and FLUX.1-dev with an $8\times$ smaller rank and up to $6.25\times$ faster quantization. Furthermore, on the Qwen3-8B LLM, rank-4 HSVDQuant improves MMLU accuracy from 61.50% to 68.17% over rank-32 SVDQuant. Overall, HSVDQuant achieves better W4A4 performance with substantially lower rank and quantization cost.

A fundamental question in neural network architecture optimization is whether the minimum hidden-neuron count required to approximate a target function within a prescribed tolerance can be computed efficiently. This paper resolves this question for two-hidden-layer ReLU networks under an $L^p(\mathbb{R}^d,\mathbb{R}^m)$ approximation constraint. For every fixed $d \ge 1$, $m \ge 1$, and $1 \le p < \infty$, we prove that computing the optimum exactly is NP-hard. The result holds even when the target is represented by a rational ReLU network whose realization is nonzero, componentwise nonnegative, compactly supported, globally Lipschitz, and continuous piecewise affine. The polynomial-time reduction from 3-SAT produces an architecture gap in which unsatisfiable formulas yield an optimum of zero, whereas satisfiable formulas yield an optimum of at least $d+2$. The proof constructs compactly supported polyhedral frustum functions realized by two-hidden-layer ReLU networks and establishes the $L^p$-density of finite linear combinations of box-frustum functions. The results offer theoretical justification for employing heuristic approximation methods in the design of ReLU neural networks, illustrating that attaining a minimal configuration within polynomial time is computationally unachievable.

While 3D spatial reasoning in dynamic egocentric environments is crucial for embodied intelligence, audio-visual large language models (AV-LLMs) lack explicit mechanisms to process and internalize global geometry directly from raw sensory streams. Existing approaches either require costly fine-tuning or underutilize the model's cross-modal reasoning capacities. In this paper, we propose FloorSAV, a novel framework that explicitly grounds spatial audio-visual context by rendering a dynamic 2D floormap. By integrating 3D point clouds, camera trajectories, spatial audio cues, and semantically grounded object landmarks, we inject this floormap into the AV-LLM as a synchronized stream with an egocentric video. AV-LLMs utilize their multi-modal capabilities to jointly reason over visual, auditory, and geometric cues in a single inference with floormap interpretation guidance. We further introduce SAVED-Bench (Spatial Audio-Visual Egocentric Benchmark with Dynamic Agents), constructing essential tasks of spatial capability in real-world scenarios: dynamic relativity, regional, and path reasoning QAs. FloorSAV improves AV-LLMs' spatial reasoning on various tasks from both SAVED-Bench and SAVVY-Bench. Studies with ground-truth floormaps demonstrate the substantial potential of FloorSAV with accurate spatial information.

Different optimizers can fit the same training data while selecting classifiers with substantially different geometries, but whether this difference provably affects population performance remains unclear. We show that row-wise normalization can achieve strictly higher population accuracy than full-batch Adam, a proxy for random-reshuffling Adam, and exact-SVD Muon in high-dimensional multiclass classification. Under an isotropic Gaussian-cloud data model, this advantage arises because row normalization's class-wise Euclidean geometry asymptotically preserves the population decision-boundary directions, whereas Adam's coordinate-wise geometry and Muon's spectral geometry introduce nonvanishing distortions. Beyond isotropy, the advantage persists for full-batch training on class means with independently oriented class-mean and test-noise covariances. It holds for power-law spectra with class-mean exponent below one, even under heavily anisotropic test noise. When both covariances are diagonal and sufficiently close, the advantage over Adam can reverse, while applying the same random rotation to both restores it by changing only their alignment with Adam's coordinate axes. Synthetic and last-layer language-model experiments support the predicted advantage.

Conditional diffusion models generate diverse, novel, and high-quality samples under prescribed conditions. However, theoretical understanding of their memorization and generalization remains limited, while recent works have characterized these behaviors primarily in unconditional settings. In this work, we analyze a random-feature conditional score model in the high-dimensional proportional limit, deriving asymptotic expressions for training and test losses. By decomposing the test loss, we show that in the overparameterized regime, increasing model width improves prediction of the condition-dependent mean while reducing within-condition prediction variance, a phenomenon we term "malign generalization." Furthermore, analyzing the training loss reveals that more informative conditions lead to memorization of training samples at smaller widths. These theoretical findings are supported by experiments with U-Net architectures on realistic data.

Humanoid loco-manipulation requires coordinated locomotion and manipulation informed by future scene evolution and whole-body motion, yet learning these capabilities is constrained by scarce robot demonstrations. Human video and motion datasets offer scalable supervision, but many contain only video or motion rather than paired video-motion data. Moreover, human motion does not directly specify executable robot actions. We present Being-M0.7, a latent world-action model that transfers visual-motion priors learned from mixed-modality human data to humanoid control through pre-training, robot mid-training, and action post-training. We curate a corpus from more than 10,000 hours of raw human-centric data, integrating video-only, motion-only, and paired video-motion streams to learn complementary visual dynamics and whole-body kinematic structure. Joint prediction of future latent visual states and motion encourages visual representations to encode future kinematics. Robot mid-training adapts this coarse-grained prior to robot viewpoints and body dynamics. During action post-training, an action expert combines visual predictive representations from the frozen, adapted prior with current images and proprioception through gated cross-attention, grounding predictive context in executable whole-body commands. Being-M0.7 achieves the highest aggregate success rate among the compared baselines on SIMPLE and matches the strongest baseline on real-world Unitree G1 loco-manipulation tasks.

Large language models (LLMs) are commonly post-trained against LLM judges and other cheap surrogates because the true reward, such as human preference, is too expensive to query at scale. This practice often leads to reward hacking, where reinforcement learning against a miscalibrated surrogate leads to undesirable side effects. In this work, we study a setting in which a small number $n$ of model outputs are annotated with ground-truth labels (e.g., from expert review) and used to recalibrate the LLM judge before optimizing against it. Prior approaches to judge recalibration are costly or heuristic, and it is known that on-policy sampling fails when the surrogate is miscalibrated on a rare set of outputs. In this work, we propose envelope sampling, a theoretically-grounded method for judge recalibration that seeks to minimize an upper bound on the regret of the post-trained model under the assumption that the human reward and re-calibrated reward lie in an $L^2$ ball around the judge. We give practical algorithms to sample from the envelope by rejection or by fine-tuning against a modified reward, and experiments on clinical note generation and on a controlled sycophancy task show that recalibrating on envelope samples mitigates reward hacking where recalibrating on base-model samples does not.

Phoneme-level models transcribe or generate speech as a sequence of phonemes, the smallest sound units that distinguish words. These models enable fine-grained pronunciation control and understanding, yet often fail on input that does not match any single training language, such as speech alternating between two languages, known as code-switching, or low-resource languages absent from training. We identify a systematic failure mode behind this, phonological interference: models assume the input is in a single language and impose its phonology, overriding local phoneme-level decisions that conflict with the assumed language. We measure interference by how often a model retains phonemes that one language has but the other lacks. On code-switched input, two phone recognizers (speech-to-phoneme models) and a phoneme-conditioned text-to-speech model lose 32% to 79% of these phonemes, but lose far fewer of the phonemes both languages share. On unseen languages, we find that phone recognizers impose the phonology of the training language they assign to the speech, and the more confident the assignment, the more they lose phonemes the unseen language has but the assigned language lacks. We probe the models' language estimate from their internal activations, and trace interference to a low dimensional subspace. On monolingual speech, steering this subspace toward another language makes the model lose the phonemes that only the original language uses and produce phonemes that only the target language has. We introduce windowed language estimation (WLE), an inference time repair that replaces the model's language estimate in this subspace with one computed from a short window around each position. On code-switched input, WLE removes 34% to 69% of the interference in all three models, and in the recognizers it leaves monolingual performance essentially unchanged.

Personalized conversational AI relies on long-term memory systems that extract facts from user utterances and store them in persistent vector stores. Despite progress in retrieval, deduplication, and lifecycle management, the formation stage, the moment a fact is first written to storage has received almost no principled attention. We identify this as the binding constraint on memory quality in production systems. Critical contextual signals, such as the distinction between a permanent user attribute and a transient situation, exist only in the original utterance and are irreversibly lost the moment extraction produces a subject-relation-object triple. No downstream process can recover them. We propose Gated Memory, a lightweight, modular formation framework that interposes two decision checkpoints between conversation and storage: an admission gate that evaluates every candidate fact against the full utterance context before extraction runs, and a conditional enrichment stage that grounds admitted facts through an entity scope taxonomy with privacy constraints. The gate evaluates only the current exchange while using prior turns as read-only reference context, and produces a structured formation record. Admitted content is decomposed into atomic facts, each categorized, tagged with provenance (directly stated versus inferred), scoped to its condition of applicability, and grounded in resolved time and place, subject to a constraint that no entity absent from the context may be asserted. On the LoCoMo-10 benchmark with atypical emotional density in utterance data, Gated Memory achieves an overall +2.6% relative improvement in LLM-judge accuracy over a strong baseline with identical retrieval and generation, establishing formation quality as a measurable constraint on memory performance.

Variational quantum algorithms are often limited by barren plateaus: gradients vanish as circuit size and depth increase, making quantum neural networks difficult to train. We propose Q-Capsule, a localized capsule-based quantum neural architecture that mitigates this problem through register partitioning, local readout, sparse inter-capsule coupling, trainable data re-uploading, and Quantum Fisher Information Matrix (QFIM)-guided adaptive depth growth. By restricting the dominant support of each observable to a small capsule and controlling inter-capsule entanglement, Q-Capsule preserves useful gradient signals while retaining communication between local quantum representations. As the register width increases, Q-Capsule consistently maintains stable gradient variance, whereas globally entangling baselines exhibit exponential suppression with a log-gradient-variance slope near -ln 2 per qubit. Q-Capsule also produces more structured optimization landscapes, higher parameter efficiency, improved robustness to depolarizing noise, and lower measurement requirements. Its adaptive policy achieves 98.1% accuracy on binary classification and 97.7% on four-class classification, while using approximately 73% fewer two-qubit gates than the fixed-deep model on the multiclass task.

Learned representations can lose latent degrees of freedom successively, suggesting a cascade of transitions whose underlying stability principle remains unclear. Here we formulate dimension-wise posterior collapse in variational autoencoder (VAE) as a fluctuation theory around partially collapsed states. Interpreting the negative evidence lower bound as an effective free energy, its quadratic expansion defines a Gaussian theory whose Hessian acts as a mass matrix for latent fluctuations. We show that the collapsed directions form an invariant fluctuation sector and derive its exact mass spectrum in terms of a conditional residual operator. A local reactivation direction lowers the free energy when the decoder variance falls below the residual spectral upper edge, with equality marking marginality. The criterion recovers principal component thresholds in the linear Gaussian VAE limit. Viewed in reverse along continuously connected branches, the reactivation boundary provides a local criterion for successive collapse. Numerical continuation experiments show successive loss of latent dimensions near these spectral marginalities. These results support a spectral cascade interpretation governed by residual information left unexplained by the surviving representation.

This study presents the Neuro-Memory Fuzzy Inference System (NeMeFIS), a hierarchical machine learning architecture that asymmetrically models acceleration and deceleration in car following behavior by integrating five human memory types procedural, working, episodic, semantic, and declarative. By linking external variables to memory functions via metaheuristics and validating them through factor and p-value analyses, NeMeFIS uncovers latent cognitive influences across Arterial, Collector, and Rural Highway corridors for different types of vehicles. Results from 54 different trained models emphasize cognitive thresholds shaped by driver perception limits and cognitive load. The trained NeMeFIS models outperform traditional statistical and conventional machine learning models in replicating realistic driving behavior, including comparisons with Linear Regression, ANFIS, and LSTM architectures. Fuzzy rule analysis reveals that declarative memory demands the highest rule, especially during deceleration, indicating complex braking decisions. Procedural memory drives acceleration, while semantic and declarative memory guide deceleration. Risk perception also emerges as a key factor, particularly on urban roads. Validated on both heterogeneous and homogeneous datasets, NeMeFIS offers a robust framework for modeling driver cognition. The findings support psychotherapeutic applications and the development of adaptive, human-like decision systems in Connected and Autonomous Vehicles (CAVs) to enhance traffic safety.

Vision-language models (VLMs) have demonstrated impressive capabilities but suffer from substantial computational overhead, as vision tokens dominate the input sequence. This motivates vision token compression as a key direction to alleviate the burden. However, with the emergence of hybrid architectures incorporating linear attention (\eg, Qwen3.5), prior methods designed for softmax attention struggle to generalize. Our analysis reveals that both attention- and similarity-based approaches suffer notable performance degradation, underscoring the urgent need for compression methods tailored to this regime. To this end, we propose V-CoLA, an efficient training-free token compression framework specifically designed for linear attention. V-CoLA introduces a novel uniqueness-aware importance criterion for identifying critical vision tokens, coupled with an adaptive token merging strategy that performs compression. All components are optimized at the implementation level to remain compatible with the chunk-wise parallelism of linear attention, ensuring strong practical value. Extensive experiments across multiple benchmarks demonstrate the superiority of V-CoLA: it achieves 99.5% of the original performance with only 50.0% of vision tokens, and over 88.0% with as few as 12.5%, while delivering a 1.86$\times$ to 6.15$\times$ prefill speedup.

On-policy distillation (OPD) enables effective capability transfer between language models, yet the mechanisms underlying its failures are not fully understood. Across code generation and mathematical reasoning, OPD with larger-scale teachers exhibits early loss plateaus, with an average final loss reduction of 25.1% after 200 updates, compared with 96.2% for self-RL teachers, obtained by further reinforcement learning (RL) training of the initial student. To understand this difference, we analyze OPD as an idealized continuous-time dynamical system in the small-learning-rate limit. Our training-log diagnostics associate these plateaus with an early decline in a gradient-based learning-signal proxy while substantial loss remains; these measurements do not establish why the underlying gradient weakens. We further prove a local recovery guarantee for teachers sufficiently close to the initial student in a shared parameterization under regularity conditions, offering a conditional explanation for the success of self-RL teachers in our experiments. Across runs with and without loss plateaus, we observe small relative parameter changes (0.025-0.098%) and high similarity between the student's representations before and after OPD (linear CKA $>0.98$ across layers). These observations suggest that limited representation adaptation may contribute to learning-signal collapse, a hypothesis that remains to be tested. Code is available at https://github.com/leizhao7/opd-learning-signals.

KV-cache entries are stored before their future queries are known, but each decoding query needs precision in different places. We study this mismatch using separate budgets for retained bits and bits fetched per query. ReadKV stores each key and value in a progressive code whose prefixes support different reconstruction precisions. For each query, it allocates key-channel prefixes using the query, computes attention from the reconstructed keys, and then allocates value-token prefixes using that attention. Stored entries remain unchanged. Each stage optimizes a calibrated distortion objective under a fixed budget; we prove exact allocation under diminishing refinement gains and relate these objectives to attention-output error. We also exhibit a finite-dimensional attention family where query-dependent access strictly outperforms every query-independent reader at the same read budget, even with unrestricted competing encoders and decoders. Across six base models, reading four bits on average from an eight-bit cache increases C4 perplexity by at most 0.66%, using about one quarter of the logical reads and half the retained capacity of a 16-bit cache. It is consistently more accurate than storing and fully reading four bits at the same payload-read budget. Retaining more bits than each query fetches is aimed at long-context decoding, where the cache bytes moved per step, rather than the weights, dominate cost. Long-context question answering and retrieval on two instruction-tuned models provide additional quality evidence. On the tested 8K-token, batch-one, single-layer workload on an NVIDIA A10G, a restricted eight-bit ReadKV reader with a two-bit mean payload-read budget has 39% lower latency than the tested TurboQuant codec.

Low-cost sensors enable scalable indoor air quality monitoring but require calibration because of nonlinear distortions, noise, and temporal drift. The conventional strict pairwise calibration setting requires a co-located reference sensor at each deployment location and does not account for spatial and temporal heterogeneity. To address these limitations, we introduce a six-month dataset comprising multivariate indoor air-quality measurements from low-cost and reference sensors with contextual metadata collected at five locations. Using this dataset, we define four evaluation scenarios. The reference-efficient and location-transfer scenarios evaluate spatial generalization, whereas the long-term drift and event-conditioned scenarios assess robustness to gradual and abrupt distribution shifts. Based on these scenarios, we derive design requirements and propose a lightweight temporal model that combines input-window compression with residual temporal and feature fusion. Experiments show strong calibration performance across all four scenarios with low edge-inference cost.

Time series forecasting aims to predict future values from historical observations and auxiliary features. We propose SteerCast, a retrieval-based latent steering method that improves decoder-only forecaster at inference time, without updating its parameters. SteerCast constructs a database from the training set by storing a representation of each history window together with a steering vector computed in the forecaster's latent space, defined as the difference between representations induced by the ground-truth continuation and by the model's own prediction. At test time, SteerCast retrieves nearest neighbors for a query history, aggregates their steering vectors, and injects the resulting signal into the forecaster's hidden states at every step of autoregressive generation, guiding predictions toward trajectories consistent with similar training cases. Experiments across diverse multivariate benchmarks and multiple horizons show that SteerCast consistently improves forecasting accuracy over the fine-tuned backbone and retrieval-based baselines, while requiring no additional training beyond the original fine-tuning and using only the training set as a retrieval corpus.

Multimodal Large Language Models (MLLMs) have demonstrated strong potential for sequential recommendation through their ability to reason over complex multimodal data. However, existing approaches either rely solely on the target user's own interaction history, neglecting collaborative signals from neighboring users, or incur substantial computational overhead through repeated MLLM inference over long interaction histories. To address these challenges, we propose MGRASRec, a multimodal graph retrieval-augmented framework for sequential recommendation. MGRASRec injects collaborative filtering signals conditioned on the candidate item directly into the MLLM prompt by retrieving structured paths from a user-item interaction graph, extended via multimodal similarity to increase coverage beyond exact co-interaction overlap. This retrieval also surfaces the history items most relevant to the candidate at no additional cost, removing the need for recurrent summarization and keeping inference to a single forward pass per candidate. All components are unified into an augmented prompt for parameter-efficient fine-tuning of an MLLM. Extensive evaluations across three publicly available datasets validate the effectiveness of MGRASRec, achieving the best performance on all metrics with particularly strong gains in ranking quality.

Weight-only post-training quantization (PTQ) relies heavily on reconstruction loss minimization to preserve model quality at low precision. We show that the weights favored by minimizing this loss need not yield better model performance on new tasks. In fact, we find that lower reconstruction loss can even degrade model performance on the same calibration data. Our analysis further shows that weights with lower reconstruction loss on calibration data can have higher loss than other weights when the distribution of input activations changes. Motivated by these observations and our analysis, we propose Distributionally Robust Quantization (DRQ), a post-hoc refinement process that minimizes worst-case reconstruction loss over a constrained set of input activation distributions. DRQ refines the integer codes representing quantized weights within the existing quantization grid, keeping quantization parameters and inference operators unchanged. Extensive experiments show that DRQ improves models quantized by six representative PTQ methods, including AWQ, GPTQ, and ParoQuant, and delivers gains across both dense and mixture-of-experts large language models. These results establish DRQ as a general post-hoc refinement framework for weight-only PTQ, achieving better downstream performance without adding inference overhead.

Preclinical models poorly predict human drug efficacy, particularly in neurological disorders. Neural activity offers a uniquely rich source of translational information because it captures high-dimensional variation in nervous-system function that can be measured in both animals and humans. However, its high dimensionality makes it difficult to distinguish conserved disease-related features from variation arising from species, recording modality and experimental context. Here, we test whether shared neural dynamics can be identified directly from electrophysiology data by learning representations organized by biological state rather than species. We develop a dual-rule contrastive learning framework that aligns corresponding mouse and human states while preserving separation between distinct phenotypes. This framework recovered conserved sensory-response structure across species and, in epilepsy, resolved distinct relationships between three mouse models and heterogeneous human patient populations. When treated animals were projected into a frozen cross-species representation, drug-induced movement towards the human-aligned healthy state retrospectively tracked known clinical efficacy across ten model-drug combinations including a disease-specific detrimental effect. The framework also identified shared disease-associated neural dynamics between Fmr1-knockout mice and human 16p11.2 copy-number variant carriers despite differences in genetic aetiology and recording modality. Together, these findings show the potential of cross-species neural representation learning to map heterogeneous human disease onto experimentally tractable preclinical states and assess whether interventions restore human-relevant circuit function.

Dense associative memory (DAM) provides an energy-based framework for memory retrieval with close connections to attention mechanisms in modern artificial intelligence. Despite growing interest in differential privacy for AI, the privacy of DAM retrieval dynamics remains relatively unexplored. In this paper, we develop a differential privacy framework for log-sum-ReLU (LSR) dense associative memory, whose finite-support retrieval dynamics pose distinctive challenges for privacy-preserving computation. We propose the Boundary-Responsive Adaptive Correction Evolution (BRACE) algorithm, a differentially private retrieval mechanism for LSR-DAM that adaptively corrects boundary-sensitive perturbations to control their cumulative effect over the retrieval trajectory. In theory, we prove that our method is minimax optimal by deriving dimension-independent terminal and full-trajectory retrieval error rates, with optimal dependence on the inverse temperature and, in the growing-horizon regime, the retrieval horizon. We further establish central limit theorems that enable uncertainty quantification for private retrieval by characterizing its asymptotic distribution and the additional variability introduced by privacy. Numerical experiments compare our proposed method with baseline differential privacy approaches and evaluate its retrieval accuracy. Together, our results provide a theoretical foundation for optimal privacy-preserving retrieval and uncertainty quantification in energy-based associative memory systems.

Diffusion and autoregression (AR) have long been seen as different categories of generative models, with diffusion specialising in continuous fields and AR specialising in discrete tokens. Recent work seeks to combine the advantages of the two models, and each hybrid fixes its decoding schedule by design. In this paper, we ask whether the performance of decoding schedules of one model can be predicted before decoding at a fixed number of steps. We describe diffusion, AR, and models in between as paths on one corruption lattice, and define the cost of a schedule as the dependence its parallel steps discard. The cost shows that the fewest steps of a zero-cost schedule are set by the geometry of the data, in the same way for tokens and for continuous fields. In particular, for data that are Markov on a graph and dependent along its paths, the fewest steps equal the graph's treedepth, which is logarithmic in the length of a sequence and linear in the side length of a grid. With fewer steps than the treedepth, every schedule pays a positive cost, whose ranking we predict before decoding with a kernel of pairwise dependence estimated from pretrained weights. Across text generation, image generation, and video generation, we verify most of the predictions about the rankings of different schedules under different metrics and benchmarks. This work therefore provides a design principle for decoding for future AR models, diffusion models, and anything in between. Our code is available at https://github.com/TSUITUENYUE/The-Lattice-of-Transition-Laws.

KV-cache quantization and linear attention are two representative approaches to tackling the storage and computational costs of Transformers. KV-cache quantization compresses individual KV entries into discrete codes but retains all entries, whereas linear attention recurrently aggregates multiple historical KV contributions into a fixed-size continuous state but can introduce interference. This contrast raises the question of whether per-KV compression and multi-KV aggregation can be bridged within a single mechanism for efficient attention. We identify RAM-Net as such a bridge through soft assignments over a discrete address space. These assignments determine recurrent updates to the continuous slot state associated with each address. Under a restricted RAM-Net construction, we prove that soft address assignments extend hard quantized matching to a separable read-write overlap that locally approximates full-attention similarity and supports recurrent aggregation. These connections further enable Transformer-to-RAM-Net weight migration through a new path based on a soft-quantized intermediate construction. Across nine pretrained Transformer models from 0.3B to 7B parameters, RAM-Net recovers an average of 87.1% of the teachers' accuracy gains over random guessing across six commonsense and knowledge tasks using only a 500M-token budget per model.

Flat loss landscapes have long been linked to better generalization in neural networks. However, its role as a causal mechanism for generalization is less established. Grokking provides an unique testbed to understand this distinction: models are prone to fit observed data using non-generalizing structure and remain in that regime for prolonged periods, transitioning to generalization only under particular training conditions. In this work, we study whether flat loss landscapes can act as a driving mechanism in this transition. While recent work has argued for flatness as a necessary geometric condition for this transition, we find that biasing training toward flatter solutions using sharpness-aware minimization (SAM) is insufficient to reliably induce this transition, despite producing flatter solutions. However, when SAM is paired with mechanisms that drive generalization such as weight decay, an interesting property emerges: SAM can accelerate the transition to generalizing solutions by up to 4x at the epoch-level. We theoretically untangle this relationship between SAM and weight decay using a minimal interpolating two-layer ReLU model with both memorizing and generalizing solutions. We show that even in this simple setup, flatness alone cannot distinguish a memorizing solution from a generalizing one, while weight decay favors generalizing solutions. However, under a local stability analysis, there exists a window where a memorizing interpolant is locally stable under gradient descent but unstable under SAM in the low-norm regime, which can explain SAM's ability to accelerate this transition. Overall, our results provide a more interpretable account of the role of flatness in driving generalization, especially in settings where models are vulnerable to minimizing loss through learning non-generalizing structure.

Representation structure is crucial for understanding Vision Transformer (ViT) architectures and their generalization behavior. However, prior studies neither isolate nor analyze module-level features nor investigate how their interactions contribute to performance estimation. In this work, we conduct the first rigorous analysis of feature information across diverse architectural scales, empirically uncover the relationship between ViT representation and generalization behavior, and leverage these insights to guide efficient ViT design. Our contributions are fivefold: Across diverse architectural scales, 1) We identify feature collapse at initialization, which leads to redundancy, and propose a reduction scheme to mitigate this issue. 2) We quantify feature information using entropy and the minimum eigenvalue, demonstrating that these metrics serve as reliable indicators for generalization prediction. 3) We show that feature in the token space provides a more faithful representation than those in embedding space. 4) We discover an unexpected finding: features produced by linear submodules within ViT layers are critical for the prediction of generalization performance. 5) Our proposed proxy improves the correlation ranking by 18-48% over prior baselines and can effectively identify ViT architectures that achieve higher accuracy at lower or comparable computational cost.