Large Models

Jailbreak attacks on large models have drawn growing attention due to their close ties to societal safety. This work identifies a practical yet unexplored jailbreak scenario, the wide-net-casting scenario, where an adversary can query a group of large models instead of a single one to elicit harmful outputs. Our analysis reveals substantial yet previously overlooked safety risks under this scenario. As a key part of our analysis, we further develop a novel jailbreak method tailored to the wide-net-casting scenario. With this tailored method, the jailbreak success rate can even reach 100% in some experiments when targeting the large models without additional safeguards, exposing wide-net-casting as a distinct, high-risk scenario that warrants attention in future evaluation and defense research.

Large language model (LLM) training today runs on clusters spanning thousands of GPUs. While this scale enables rapid model advances, developing, debugging, and performance-tuning the training framework inevitably becomes complex and costly. This is because engineers often need to reproduce production behaviors to diagnose failures or evaluate optimizations, thereby demanding frequent and even exclusive access to production-scale clusters -- which becomes increasingly hard given that the majority of GPUs are already committed to production workloads. Simulation relies on complex performance models that are difficult to maintain, and downscaled experiments often fail to capture scale-dependent behaviors. We present PrismLLM to decouple large-scale execution from the need to access large clusters, enabling engineers to run and observe ranks of interest under faithful large-scale behavior using only a few GPUs. PrismLLM constructs a high-fidelity execution graph via a slicing-based approach that captures computation, communication, and dependencies of the target scale. Then, PrismLLM performs hybrid emulation where selected ranks execute the original program while the remaining ranks are replayed as virtual participants. Experiments on large-scale LLM training workloads show that PrismLLM accurately reproduces performance and memory behavior, achieving only 0.58% average error in iteration time and less than 0.01% error in peak GPU memory usage. PrismLLM can emulate clusters of up to 8192 GPUs using fewer than 1% of the physical GPUs required by the original deployment.

The dynamic range of activations is a first-order constraint for low-bit quantization, activation scaling, and stable LLM inference. Prior work characterized outlier features and massive activations on pre-2024 LLaMA-style models, and the downstream activation-quantization stack inherits that picture without revisiting it for the post-LLaMA open-model boom. We ask the deployment-oriented question: how large can activations get in modern open LLMs, and how does this magnitude vary across families, generations, and training stages? Under a unified pipeline (5,000-sample multi-domain corpus, family-specific tokenization, identical hooks across embeddings, hidden states, attention, MLP/MoE, SwiGLU gates, and final norm), we measure global and layerwise maxima on 27 checkpoints from 8 open families spanning dense, MoE, vision-language, intermediate-training, and instruction-tuned variants. We find that (i) global maxima span over nearly four orders of magnitude at comparable parameter counts, with Qwen3.5 and MoE checkpoints in the 10^2 to 10^3 range and Gemma3-27B-it reaching ~7 x 10^5; (ii) cross-family and cross-generation comparisons break simple monotonic scaling; and (iii) MoE checkpoints exhibit 14.0-23.4x lower peaks than matched-scale dense counterparts, while the residual stream carries the global maximum in 22/24 checkpoints. A lightweight INT-8 sanity check shows that measured maxima co-vary with low-bit reconstruction error via activation-scale selection. We conclude that maximum activation magnitude is a model property tied to family, architecture, and training stage - not a simple byproduct of size - and should be measured and reported alongside any open-weight release before low-bit deployment. The code is publicly available at https://github.com/clx1415926/Max_act_llm.

Despite the remarkable success of transformer architectures in natural language processing, their scalability limitations remain poorly understood through systematic empirical analysis. This paper presents the first comprehensive large-scale evaluation of 118 transformer models across seven distinct architectural categories, revealing fundamental performance walls that manifest as hard deployment constraints. Our systematic benchmarking methodology uncovers a critical scalability crisis: while 88.1% of models successfully process sequences up to 512 tokens, this drops dramatically to 44.9% at 1024 tokens, with complete failure (0%) at 2048 tokens. Through rigorous analysis of loading times, memory consumption, and computational efficiency across sequence lengths from 128 to 2048 tokens, we demonstrate that compressed models achieve superior parameter efficiency (649.2 tokens/sec/M parameters) compared to large generative models (12.5 tokens/sec/M). Our findings challenge prevailing scaling assumptions and provide the first quantitative evidence that the theoretical O(n2) attention complexity translates into measurable performance walls. This work establishes new benchmarking methodologies for transformer evaluation and provides critical insights for practical deployment decisions in production environments.

Scaling LLMs requires tremendous computational resources, and recent advances in AI have gone hand in hand with massive amounts of capital expenditure. While it is established that scaling up LLMs reliably increases model quality (quantified in terms of loss or downstream evaluations), it is unclear how these quality improvements translate to potential revenue, and whether revenue increases would offset costs of larger-scale training and inference. In this work, we develop an economic model for characterizing the rational behavior of an LLM training firm by combining scaling laws with microeconomic theory. Under our model of firm behavior, LLM quality can be increased with more parameters and training tokens, leading to more potential adoption by consumers, who each have a quality threshold for using the LLM. On the other hand, additional parameters and training tokens both incur additional costs. We analyze the profit maximization problem for this model under compute-bound and data-bound regimes. In the compute-bound regime, optimal model size and token budget track hardware efficiency $E$ (FLOPs/$) at a near-linear rate; total training cost then scales sub-quadratically in $E$. Data efficiency improvements incentivize larger models and training expenditure. When we are limited to $D$ data, profit-optimal training expenditure scales as $D^2/E$, i.e, increase with data and decreases with hardware efficiency (as well as data efficiency). Finally, we analyze practical trends in training expenditure: current trends are consistent with our most permissive model variants in the compute-bound regime, but are not profit-optimal in the data-bound regime or assuming hardware advances will stall. Overall, our results provide a theory of profit-optimal LLM training, providing a foundation for engaging critically with industry statements and supporting long-term economic decision making.

Hyperparameter transfer across model architectures dramatically reduces the amount of compute necessary for tuning large language models (LLMs). The maximal update parameterization (μP) ensures transfer through principled mathematical analysis but can be challenging to derive for new model architectures. Building on the spectral feature-learning view of Yang et al. (2023a), we make two advances. First, we promote spectral norm conditions on the weights from a heuristic to the definition of feature learning, and as a consequence arrive at the Complete-P depth and weight-decay scalings without recourse to lazy-learning. Second, we consider a modified spectral norm that preserves the valid scaling law of network weights when weight matrices are not full rank. This enables (to our knowledge, the first) derivation of μP scalings for grouped-query attention (GQA). We demonstrate the efficacy of our theoretical derivations by showing learning rate transfer across the GQA repetition hyperparameter as well as experiments regarding transfer over weight decay.

Scaled Outer Product (SOP) is a post-training quantization methodology for large language model weights, designed to deliver near-lossless fidelity at 4.5--6 bits per weight on hardware with per-layer LUT decode. The methodology combines per-layer search of fixed and dynamic codebook pairs selected by a per-block selection bit, signed per-block scales, activation-weighted cosine selection, and multiple-choice knapsack promotion of sensitive layers with outlier and sparse-residual correction. Fixed codebooks include NF4, BOF4, Split87, and SH4; per-layer optimized codebooks (DD4) are hosted in LUT SRAM. A new hardware-efficient LUT output format (HIF) is proposed to improve performance, energy, and cost. Across six open model families, the recommended FP6 operating point (E2M3sUE4M4, 6.5 bpw) achieves lower weight reconstruction error than the conventional per-layer-POT FP8 baseline (E4M3, 8.0 bpw) at 1.5 bpw lower storage cost, demonstrating that block-scaled small atoms with carefully chosen scale precision can replace conventionally-deployed FP8. Full evaluation across the 4.5--6 bpw range, including layer promotion and sparse residual correction, is reported in a companion paper.

Recent frontier large language models predominantly rely on Mixture-of-Experts (MoE) architectures. Despite empirical progress, there is still no principled understanding of how hyperparameters should scale with network width $N$, expert width $N_e$, number of experts $M$, sparsity $K$, and depth $L$ to ensure both stability and optimal performance at scale. We take a principled step toward resolving this gap by analyzing three different scaling regimes: (I) co-scaling $N\asymp N_e$, (II) co-scaling $N\asymp M\asymp K$, and (III) full proportional scaling of $N, N_e, M$, and $K$. For each regime, we develop a novel Dynamical Mean Field Theory (DMFT) description of the limiting training dynamics of MoEs that provides a formal foundation for our analysis. Within this framework, we derive the unique parameterization for SGD and Adam satisfying all maximal-update ($μ$) desiderata. We then show that the resulting $μ$P prescription does not reliably induce monotonic improvement with scale or robust learning-rate transfer. We trace these pathologies to scale-dependent observables in the aggregation dynamics, which motivates a refined set of desiderata that we term maximal scale stability. Guided by this principle, we derive a Maximally Scale-Stable Parameterization (MSSP) for both SGD and Adam in all three scaling regimes, and characterize the corresponding limiting dynamics - qualitatively distinct from the $μ$P limit - through a separate DMFT analysis. Experiments verify that MSSP robustly recovers learning rate transfer and monotonic improvement with scale across regimes. Combined with existing depth-scaling theory, these results provide a complete scaling prescription for MoE architectures as a function of width, depth, expert width, and number of experts.

Generative recommendation (GR) has emerged as a promising paradigm that replaces fragmented, scenario-specific architectures with unified Transformer-based models, exhibiting scaling-law behavior where recommendation quality improves systematically with increased model capacity and training data. However, deploying GR at scale on Ascend NPUs faces fundamental system-level challenges. These challenges are further exacerbated on Ascend NPUs due to the absence of high-performance implementations for jagged operators and the architectural mismatch between irregular sparse primitives and NPU's dense-computation-optimized design. In this paper, we present \model, an Ascend-affinity training system for generative recommendation that systematically addresses these bottlenecks through three core innovations: (i) Ascend-affinity jagged acceleration, including fusion operators that eliminate padding redundancy and dynamic load balancing that reduces inter-device imbalance from 47% to 2.4%; (ii) distributed communication optimization, comprising hierarchical sparse parallelism, semi-asynchronous training with proven convergence guarantees, and fine-grained pipeline orchestration that sustains 94% NPU utilization; and (iii) negative sampling optimization via asynchronous offloading, jaggedness-aware FP16 quantization, and intra-batch logit sharing that expand the effective negative space without additional embedding lookups. Evaluated on the KuaiRand-27K dataset, \model supports training at up to 0.2B parameters and achieves 54.71% MFU with near-linear scalability (0.97).

Model growth from a given checkpoint aims to accelerate training of a larger model, offering potential resource savings. Despite recent interest, warmstarting has seen limited practical adoption in large-scale training. We attribute this to two underexplored factors: (1) an overemphasis on preserving the smaller model's performance at initialization, which constrains operator design for new architectures, and (2) insufficient analysis of how growth interacts with hyperparameters and scaling behavior, compounded by inconsistent growth factors across the literature. We show that preserving the base model's initial post-growth performance is not necessary for strong final performance, and that simple, architecture-agnostic growth strategies can outperform more complex warmstarting operators. Crucially, we empirically identify an upper bound on the growth factor $g$ beyond which training from scratch is more efficient. We observe this across multiple ablation setups. Notably, this limit is also present, but unreported, in prior published results. Across our experiments on dense MLPs and dense language models, we find that a $2\times$ growth factor is the most reliable in yielding convergence speedups, with gains most pronounced under 20 tokens/parameter budgets and diminishing as budget increases. We fit scaling laws over these observations to provide predictive guidance for practitioners deciding when and how much to grow. Together, our analysis provides practical guidelines and empirical limits for model growth.

Sparse Mixture-of-Experts (MoE) models offer a powerful way to scale model size without increasing compute, as per-token FLOPs depend only on k active experts rather than the total pool of E experts. Yet, this asymmetry creates an MoE efficiency paradox in practice: adding more experts balloons memory and communication costs, making actual training inefficient. We argue that this bottleneck arises in part because current MoE training allocates too many experts from the beginning, even though early-stage data may not fully utilize such capacity. Motivated by this, we propose EMO, a simple progressive training framework that treats MoE capacity as expandable memory and grows the expert pool over the course of training. EMO explicitly models sparsity in scaling law to derive stage-wise compute-optimal token budgets for progressive expansion. Empirical results show that EMO matches the performance of a fixed-expert setup in large-scale experiments while improving wall-clock efficiency. It offers a surprisingly simple yet effective path to scalable MoE training, preserving the benefits of large expert pools while reducing both training time and GPU cost.

Scaling laws predict loss from compute but not how capabilities interact. We measure the coupling between reasoning and truthfulness across 63 base models from 16 families and find a regime change invisible to loss curves: below a family-dependent critical scale N_c, capabilities anticorrelate (r = -0.989, p = 4 x 10^{-5} nonparametric permutation test); above it, they cooperate. N_c ~ 3.5B parameters [2.9B, 13.4B] (bootstrap 95% CI), but model size is not the only variable that determines phase. Architecture, data curation, and training recipe each shift N_c independently: curated training eliminated the coupling dip between Qwen generations (0.025 to 0.830 at matched scale), Gemma-4 at 4B achieves coupling 0.871, characteristic of 13B+ standard-trained models, through distillation and architectural innovation, and Phi at 1B matches web-trained coupling at 10B through data curation alone. Width normalization eliminates the anticorrelation across all tested families, supporting an output-projection bottleneck. Internally, 38 of 40 models show zero competing attention heads. A sparse-regression ODE cross-predicts held-out Llama-2 at 5.6% error. The diagnostic requires no model internals -- only public benchmark scores across a model family. The cooperative regime extends to the frontier (r = +0.72, 34 models, 10 labs). A proof-of-concept intervention confirms the bottleneck is exploitable: adding a single truth-direction vector at the identified layer corrects 60% of misaligned outputs in the tax phase with zero retraining -- a surgical, per-inference correction that requires no weight modification. Code, data, an open-source steering CLI for any open-weight model, and an interactive dashboard for phase diagnosis are released: https://zehenlabs.com/cape/.

We explore the effectiveness of an LLM-guided query refinement paradigm for extending the usability of embedding models to challenging zero-shot search and classification tasks. Our approach refines the embedding representation of a user query using feedback from a generative LLM on a small set of documents, enabling embeddings to adapt in real time to the target task. We conduct extensive experiments with state-of-the-art text embedding models across a diverse set of challenging search and classification benchmarks. Empirical results indicate that LLM-guided query refinement yields consistent gains across all models and datasets, with relative improvements of up to +25% in literature search, intent detection, key-point matching, and nuanced query-instruction following. The refined queries improve ranking quality and induce clearer binary separation across the corpus, enabling the embedding space to better reflect the nuanced, task-specific constraints of each ad-hoc user query. Importantly, this expands the range of practical settings in which embedding models can be effectively deployed, making them a compelling alternative when costly LLM pipelines are not viable at corpus-scale. We release our experimental code for reproducibility, at https://github.com/IBM/task-aware-embedding-refinement.

A major recent advance in quantization is given by microscaled 4-bit formats such as NVFP4 and MXFP4, quantizing values into small groups sharing a scale, assuming a fixed floating-point grid. In this paper, we study the following natural extension: assume that, for each group of values, we are free to select the "better" among two or more 4-bit grids marked by one or more bits in the scale value. We formalize the power-of-two-grids (PO2) problem, and provide theoretical results showing that practical small-group formats such as MXFP or NVFP can benefit significantly from PO2 grids, while the advantage vanishes for very large groups. On the practical side, we instantiate several grid families, including 1) PO2(NF4), which pairs the standard NF4 normal grid with a learned grid, 2) MPO2, a grid pair that is fully learned over real weights and activations, 3) PO2(Split87), an explicit-zero asymmetric grid and 4) SFP4, a TensorCore-implementable triple which pairs NVFP4 with two shifted variants. Results for post-training quantization of standard open models and pre-training of Llama-like models show that adaptive grids consistently improve accuracy vs single-grid FP4 under both weight-only and weight+activation. Source code is available at https://github.com/IST-DASLab/GridGames.

Modern large language model (LLM) training is inherently dynamic: resource fluctuations, RLHF phase shifts, and cluster elasticity continually reshape the optimal parallelism layout, posing a significant challenge to existing training frameworks built around a static execution model. We present DynaTrain, a distributed training system for sub-second, online reconfiguration across arbitrary multi-dimensional parallelism. At its core, we propose a Virtual Parameter Space (VPS) abstraction that unifies all distributed training states under one logical coordinate space, turning any parallelism configuration into a deterministic mapping and collapsing complex transition into manageable geometric intersections. On top of VPS, a state routing-and-transition layer executes rank-local transfers under a memory-aware, deadlock-free schedule, and an Elastic Device Manager overlaps new-world construction with ongoing training to mask topology-change cost. On dense and MoE models up to 235B parameters, DynaTrain reconfigures a 70B dense model in under 2s and a 235B MoE model in 4.36s, outperforming state-of-the-art checkpoint-based and elastic systems by up to three orders of magnitude while preserving correctness.

Mixture-of-Experts (MoE) architectures have become standard in large language models, yet many of their core design choices - expert count, granularity, shared experts, load balancing, token dropping - have only been studied one or two at a time over narrow configuration ranges. It remains an open question whether these choices can be optimized independently, without considering interactions. We present the first systematic study of over 2,000 pretraining runs spanning models up to 6.6B total parameters, in which we exhaustively vary total experts, expert dimension, heterogeneous expert sizing within a single layer, shared expert size and load-balancing mechanisms. We find that at every active-parameter scale that we study, performance consistently improves with total MoE parameters even at extreme active expert parameter ratios like 128.Further, the optimal expert size is nearly invariant to total parameter count and depends only on active parameter count. Third, we see that other choices like shared experts, heterogeneous experts and load-balancing settings have small effects relative to expert count and granularity, although dropless routing yields a consistent gain. Overall, our results suggest a simpler recipe: focus on expert count and granularity, other choices have minimal effect on final quality.

Test-time scaling has become an effective paradigm for improving the reasoning ability of large language models by allocating additional computation during inference. Recent structured approaches have further advanced this paradigm by organizing inference across multiple trajectories, refinement rounds, and verification-based feedback. However, existing structured test-time scaling methods either weakly coordinate parallel reasoning trajectories or rely on noisy historical information without explicitly deciding what should be retained and reused, limiting their ability to balance exploration and exploitation. In this work, we propose TMAS, a framework for scaling test-time compute via multi-agent synergy. TMAS organizes inference as a collaborative process among specialized agents, enabling structured information flow across agents, trajectories, and refinement iterations. To support effective cross-trajectory collaboration, TMAS introduces hierarchical memories: the experience bank reuses low-level reliable intermediate conclusions and local feedback, while the guideline bank records previously explored high-level strategies to steer subsequent rollouts away from redundant reasoning patterns. Furthermore, we design a hybrid reward reinforcement learning scheme tailored to TMAS, which jointly preserves basic reasoning capability, enhances experience utilization, and encourages exploration beyond previously attempted solution strategies. Extensive experiments on challenging reasoning benchmarks show that TMAS achieves stronger iterative scaling than existing test-time scaling baselines, with hybrid reward training further improving scaling effectiveness and stability across iterations. Code and data are available at https://github.com/IQuestLab/tmas.

The scaling laws guiding modern model training were calibrated for a single regime: data-rich, single-epoch pretraining. The dominant such scaling law form, Chinchilla's $L = E + A/N^α+ B/D^β$, has three structural limitations outside that regime: it diverges as unique data shrinks instead of saturating at the uninformed baseline; it cannot represent overfitting when capacity exceeds the data; and it conflates total examples seen with unique examples available. We propose a closed-form extension, $L(N, D, T) = E + (L_0 - E)\,h/(1+h)$ with $h = a/N^α+ b/T^β+ c\,N^γ/D^δ$, that decomposes loss into undercapacity, undertraining, and overfitting terms. It saturates between the irreducible loss $E$ and an uninformed baseline $L_0$ fixed by the loss type, and reduces to Chinchilla in the data-rich, single-epoch limit. We validate it on four multi-epoch experiments spanning four architecture families (MLPs, ResNets, Fourier neural operators, and transformers) across vision, scientific ML, and language domains, and refit it to five published LLM scaling-law grids. Extrapolating to higher compute and larger unique data than seen at fit time, our form achieves state-of-the-art RMSE on every published LLM grid we evaluate and on most cells of our constructed experiments. Once calibrated, the form admits a cost-aware allocation that recovers Chinchilla's optimum when data is free and shifts toward smaller corpora and more epochs as data grows expensive.

Training large language models requires optimization algorithms that are not only statistically effective, but also computationally and memory efficient at extreme scale. Although Adam remains the dominant optimizer for large-scale language-model pretraining and fine-tuning, recent work has revisited nearly every component of the optimization stack: adaptive moment estimation, decoupled weight decay, memory footprint, curvature approximation, sign-based updates, large-batch stability, low-rank gradient structure, and matrix-wise orthogonalized updates. This survey reviews optimizer design for large language models through a systems-and-optimization lens. We organize the literature into classical first-order optimizers, adaptive optimizers, memory-efficient variants, second-order and curvature-aware methods, sign-based and discovered optimizers, low-rank and projection-based methods, and matrix-based optimizers such as Muon. We also discuss benchmarking methodology, including hyperparameter fairness, scale dependence, wall-clock efficiency, token efficiency, memory overhead, and downstream evaluation. We argue that optimizer research for LLMs is entering a new phase: moving from single-algorithm speedup claims toward rigorous, scale-aware comparisons that jointly evaluate convergence, stability, memory, and implementation complexity.

Looped language models repeat a set of transformer layers through depth, reducing memory costs and providing natural early-exit points at loop boundaries. However, looped models do not scale as favorably as standard transformers with unique layers. We compare standard and Mixture-of-Experts (MoE) transformers, with and without looping, and find two main results. First, we find Looped-MoE models scale better than the standard baseline while dense looped models do not. We trace this to routing divergence between loops: in Looped-MoE models, different experts are activated on each pass through the same shared layers, recovering expressivity without additional parameters. Our second finding is that looped models have better compute-quality trade-offs with early exits than standard models. Because each loop ends with the same layers that produce the final output, loop boundaries are superior exit points, as confirmed by earlier output convergence at these points. In sum, we provide a clear direction for scaling looped models: a Looped-MoE model with early exits can not only beat standard transformers at scale, but also enable significant memory and inference savings with minimal degradation in quality.

We introduce a predictive model that estimates the pre-training loss of large models from model size (N), batch size (B) and number of weight updates (K). This is the first loss prediction model that can handle changing batch size. The model outperforms Chinchilla's loss model, a model of the test loss using the batch size and number of tokens, in terms of projecting the loss at extrapolated compute budgets (up to 1000 folds). A natural use of the model is to find optimal N, B, K configurations under explicit and compound resource constraints like time, memory and compute. In our experiments, the model-selected configurations are close to ground-truth optimal. Our work advocates for loss prediction as a better alternative to heuristic-based laws, which are growing in complexity. The implementation is available on https://github.com/chuningxdy/Noisy-Quadratic-System.

Chain-of-Thought (CoT) prompting symbolized a huge improvement of reasoning capabilities of Large Language Models (LLMs). However, scaling up test-time computation yields extensive CoT sequences, introducing severe inference latency and key-value (KV) cache memory bottlenecks. While structural pruning offers a fundamental, hardware-aware solution to alleviate static parameter burdens, existing magnitude-based methods may cut off the neurons of CoT: by over-indexing on discrete cross-entropy objectives, these heuristics fall into a \textit{magnitude trap}: they prioritize high-frequency, low-information syntactic tokens and trigger a disappointing reasoning collapse at high sparsities (e.g., 40%). To overcome this topological phase transition, we propose \textsc{Relative Kinetic Utility} (RKU), a novel theoretical framework that elevates discrete pruning to a continuous kinetic integral over the depth manifold of the model based on Alternating Gradient Flow(AGF). By modifying it with Fisher trace normalization, RKU acts as a lightweight curvature-aware normalization to isolate \textit{kinetic spikes} -- the fundamental structural pathways responsible for high-curvature logical routing. Extensive experiments on Qwen-2.5-7B and LLaMA-3-8B improves performance in the high-sparsity regime around 40%. RKU attains 13.34% accuracy on GSM8K at 40% sparsity, outperforming the strongest baseline, and appears to better preserve reasoning-relevant representations under out-of-distribution evaluation.

Existing approaches to LLM personalization focus on constructing better personalized models or inputs, while treating inference as a single-shot process. In this work, we study Test-Time Personalization (TTP) along an unexplored axis: scaling inference-time computation by sampling N candidates from a personalized policy model and selecting the best with a personalized reward model. We prove that oracle selection yields expected utility growing logarithmically with the number of sampled candidates, establishing a theoretical ceiling for test-time scaling. However, standard reward models fail to realize this potential. To diagnose why, we derive a unified scaling law that decomposes any reward model's Best-of-N curve into four measurable quantities and reveals two failure modes, user-level collapse (near-constant prediction for some users) and query-level reward hacking (negative correlation with true quality for some queries). Guided by this law, we propose a probabilistic personalized reward model whose learned variance effectively mitigates both failure modes. Experiments confirm both elements of our framework: TTP delivers consistent scaling across multiple policy models and personalized text generation tasks, and our scaling law closely matches observed scaling curves across reward-model variants.

Over the last two decades, language modeling has experienced a shift from the use of predominantly recurrent architectures that process tokens sequentially during training and inference to non-recurrent models that process sequence elements in parallel during training, which results in greater training efficiency and stability at the expense of lower inference throughput. Here we introduce the Structured Recurrent Mixer, an architecture that allows for algebraic conversion between a sequence parallel representation at train time and a recurrent representation at inference, notably without the need for specialized kernels or device-specific memory management. We show experimentally that this dual representation allows for greater training efficiency, higher input information capacity, and larger inference throughput and concurrency when compared to other linear complexity models. We postulate that recurrent models are poorly suited to extended sequence length scaling for information-rich inputs typical of language, but are well suited to scaling in the sample (batch) dimension due to their constant memory per sample. We provide Mojo/MAX inference implementations of SRMs exhibiting 12x the throughput and 170x the concurrency of similarly powerful Transformers inferenced on vLLM, increases characteristic of Pytorch implementations resulting in a 30% increase in compute-constant GSM8k Pass@k. We conclude by demonstrating that SRMs are effective reinforcement learning training candidates.

Microscaling is a critical technique for preserving the quality of Large Language Models (LLMs) quantized to ultra-low precision formats. Intuitively, finer block sizes should yield lower quantization error; however, a paradox recently identified by Fasoli et al. (2026) demonstrates that standard abs-max scaling can actually result in degraded model quality as block sizes shrink. In this work, we investigate the underlying mechanics of this phenomenon. We demonstrate that this degradation is not an inherent limitation of finer granularity, but is primarily driven by how elements in smaller blocks statistically cluster closer to their local block maximum, interacting poorly with the coarse subnormal E4M3 values used as scaling factors. Specifically, we show that i) preventing the scaling factor from underflowing to zero mitigates errors caused by extreme underflow, ii) targeted algorithmic interventions like the 4-over-6 methodology that give more flexibility to the choice of scaling factor resolve the paradox for larger values, and iii) a brute-force search establishes an optimal baseline, confirming that the theoretical Mean Squared Error (MSE) strictly improves with finer block sizes. Ultimately, our findings highlight a critical insight for hardware-software co-design: the block-size paradox is partially an artifact of naive scale selection. While using hierarchical scaling factors or wider formats like UE5M3 interchangeably resolves much of the quality loss, we found the 4-over-6 scale selection heuristic can even further improve quality, especially for very small block sizes. Consequently, maximizing the performance of next-generation ML accelerators will require treating silicon format specifications and software scaling algorithms as tightly coupled design choices.

Neural scaling laws approximate a language model's loss as a power-law function of parameter count $N$ and token count $D$. Following Chinchilla-style compute-optimal training, many studies fit scaling laws from runs performed under a fixed tokens-per-parameter (TPP) ratio $k$ and set $D = kN$. We show that this collinear design, combined with the empirically common near-equality of the exponents governing $N$ and $D$, induces an inherent ill-conditioning in the Gauss-Newton least-squares problem: the condition number of the design grows as the inverse square of the gap between the $N$ and $D$-exponents. The scale coefficients become practically unidentifiable, with confidence intervals inflating by an order of magnitude or more, yielding a ``sloppy'' model whose extrapolations degrade sharply off the training ray. We prove this for four scaling-law formalisms and derive a closed-form TPP-diversity threshold that is necessary and sufficient for well-conditioned estimation. Empirically, non-collinear designs outperform collinear ones on held-out splits with a 97.3% win rate across four laws, five corpora, multiple floating point precision modes. We further show the degeneracy is rooted in Jacobian geometry and is not an artifact of the loss function: any smooth estimation objective whose curvature involves the Jacobian inherits the same ill-conditioning.

We investigate the origins of massive activations in large language models (LLMs) and identify a specific layer named the \textbf{Massive Emergence Layer (ME Layer)}, that is consistently observed across model families, where massive activations first emerge and subsequently propagate to deeper layers through residual connections. We show that, within the ME Layer both the RMSNorm and the FFN parameters jointly contribute to the emergence of massive activations. Once formed, the massive activation token representation remains largely invariant across layers, reducing the diversity of hidden representations passed to the attention module. Motivated by this limitation, we propose a simple and effective method to reduce the rigidity of the massive activation token. Our approach consistently improves LLM performance across multiple tasks, including instruction following and math reasoning, in both training free and fine tuning settings. Moreover, we show that our method mitigates attention sinks by selectively weakening their influence, elucidating their origin at the hidden state level and shedding new light on principled mitigation strategies.

Test-time scaling (TTS) has become an effective approach for improving large language model performance by allocating additional computation during inference. However, existing TTS strategies are largely hand-crafted: researchers manually design reasoning patterns and tune heuristics by intuition, leaving much of the computation-allocation space unexplored. We propose an environment-driven framework, AutoTTS, that changes what researchers design: from individual TTS heuristics to environments where TTS strategies can be discovered automatically. The key to AutoTTS lies in environment construction: the discovery environment must make the control space tractable and provide cheap, frequent feedback for TTS search. As a concrete instantiation, we formulate width--depth TTS as controller synthesis over pre-collected reasoning trajectories and probe signals, where controllers decide when to branch, continue, probe, prune, or stop and can be evaluated cheaply without repeated LLM calls. We further introduce beta parameterization to make the search tractable and fine-grained execution trace feedback to improve discovery efficiency by helping the agent diagnose why a TTS program fails. Experiments on mathematical reasoning benchmarks show that the discovered strategies improve the overall accuracy--cost tradeoff over strong manually designed baselines. The discovered strategies generalize to held-out benchmarks and model scales, while the entire discovery costs only $39.9 and 160 minutes. Our data, and code will be open-source at https://github.com/zhengkid/AutoTTS.

In large-scale reinforcement learning (RL) systems with decoupled Trainer-Rollout execution, the Trainer must regularly synchronize policy weights to the Rollout side to limit policy staleness. When inter-node bandwidth is abundant, such synchronization is usually only a small fraction of end-to-end cost. As model size grows, however, the communication demand rises rapidly. In bandwidth-constrained or network-variable deployments -- for example, cross-datacenter or cross-cluster settings, heterogeneous resource pools, and online RL -- weight synchronization can become a dominant bottleneck for throughput and tail latency. We observe that, in mainstream large-model RL training, the locations where parameters actually change are highly sparse at the element level (often 99%+ sparsity). Building on this observation, we propose and implement SparseRL-Sync, which replaces full-weight transfers with a lossless sparse update payload (indices and values) that can be exactly reconstructed on the inference side, thereby preserving 100% fidelity. Under a simplified cost model, sparse synchronization reduces the per-update communication volume from S to approximately S/X; with 99% sparsity (X ~ 100), this yields about a 100x reduction in transmitted data. Combined with appropriate bucketing, SparseRL-Sync also reduces launch and control-plane overhead, significantly improving scalability and end-to-end efficiency in bandwidth-limited and highly asynchronous RL settings.

Beyond training-time optimization, scaling test-time computation has emerged as a key paradigm to extend the reasoning capabilities of Large Language Models (LLMs). However, most existing methods adopt a rigid Planning-before-Trial (PbT) policy, which inefficiently allocates test-time compute by incurring planning overhead even on directly solvable problems. We propose Planning-after-Trial (PaT), an adaptive policy for code generation that invokes a planner only upon verification failure. This adaptive policy naturally enables a heterogeneous model configuration: a cost-efficient model handles generation attempts, while a powerful model is reserved for targeted planning interventions. Empirically, across multiple benchmarks and model families, our approach significantly advances the cost-performance Pareto frontier. Notably, our heterogeneous configuration achieves performance comparable to a large homogeneous model while reducing inference cost by approximately 69%.

Modern Mixture-of-Experts (MoE) architectures allocate expert capacity through a rigid per-layer rule: each transformer layer owns a separate expert set. This convention couples depth scaling with linear expert-parameter growth and assumes that every layer needs isolated expert capacity. However, recent analyses and our routing probe challenge this allocation rule: replacing a deeper layer's learned top-k router with uniform random routing drops downstream accuracy by only 1.0-1.6 points across multiple production MoE models. Motivated by this redundancy, we propose UniPool, an MoE architecture that treats expert capacity as a global architectural budget by replacing per-layer expert ownership with a single shared pool accessed by independent per-layer routers. To enable stable and balanced training under sharing, we introduce a pool-level auxiliary loss that balances expert utilization across the entire pool, and adopt NormRouter to provide sparse and scale-stable routing into the shared expert pool. Across five LLaMA-architecture model scales (182M, 469M, 650M, 830M, and 978M parameters) trained on 30B tokens from the Pile, UniPool consistently improves validation loss and perplexity over the matched vanilla MoE baselines. Across these scales, UniPool reduces validation loss by up to 0.0386 relative to vanilla MoE. Beyond raw loss improvement, our results identify pool size as an explicit depth-scaling hyperparameter: reduced-pool UniPool variants using only 41.6%-66.7% of the vanilla expert-parameter budget match or outperform layer-wise MoE at the tested scales. This shows that, under a shared-pool design, expert parameters need not grow linearly with depth; they can grow sublinearly while remaining more efficient and effective than vanilla MoE. Further analysis shows that UniPool's benefits compose with finer-grained expert decomposition.

Linear Attention (LA) offers a promising paradigm for scaling large language models (LLMs) to long sequences by avoiding the quadratic complexity of self-attention. Recent LA models such as Mamba2 and GDN interpret linear recurrences as closed-form online stochastic gradient descent (SGD), but naive SGD updates suffer from rapid information decay and suboptimal convergence in optimization. While momentum-based optimizers provide a natural remedy, they pose challenges in simultaneously achieving training efficiency and effectiveness. To address this, we develop a chunkwise parallel algorithm for LA with a stepwise momentum rule by geometrically reordering the update coefficients. Further, from a dynamical systems perspective, we analyze the momentum-based recurrence as a second-order system that introduces complex conjugate eigenvalues. This analysis guides the design of stable gating constraints. The resulting model, Momentum DeltaNet (MDN), leverages Triton kernels to achieve comparable training throughput with competitive linear models such as Mamba2 and KDA. Extensive experiments on the 400M and 1.3B parameter models demonstrate consistent performance improvements over strong baselines, including Transformers, Mamba2 and GDN, across diverse downstream evaluation benchmarks. Code: https://github.com/HuuYuLong/MomentumDeltaNet .

Training loss and throughput can hide distinct internal representation in language-model training. To examine these hidden mechanics, we use spectral measurements as practical and operational diagnostics. Using a controlled family of decoder-only models adapted from the modded NanoGPT codebase, we introduce an empirical protocol based on activation covariance and per-sample gradient SVD spectra. This dual-view reveals three empirical findings and one mechanistic explanation. First, batch size acts as a latent determinant of representation geometry: runs that reach equal loss settle into systematically distinct activation spectra. Second, the activation covariance tail measured early in training reliably forecasts downstream token efficiency. Third, movement of the activation spectrum head (leading modes), together with gradient spectra, characterizes underlying learning-dynamics changes, separating learning-side architectural improvements from primarily execution-side gains. These predictive and diagnostic signals persist across the 12-, 36-, and 48-layer model tiers. Finally, a mechanistic model proves the main observations and explains how activation covariance spectra correlate with task-aligned feature learning.

Frontier models increasingly adopt Mixture-of-Experts (MoE) architectures to achieve large-model performance at reduced cost. However, training MoE models on HPC platforms is hindered by large memory footprints, frequent large-scale communication across heterogeneous networks, and severe workload imbalance. To characterize these challenges, we develop a mathematical model that quantifies memory, compute, and communication requirements for MoE configurations under various parallelization schemes, verified through micro-benchmarking, code instrumentation, and hardware profiling. Our analysis identifies performance bottlenecks: all-to-all latency at scale from expert parallelism, insufficient compute-communication overlap, low GPU utilization from imbalanced skinny GEMMs, and the absence of platform-aware hybrid parallelization strategies. To address these, we introduce Piper, a framework that leverages resource modeling to identify efficient training strategies for MoE models on target HPC platforms, applying pipeline parallelism with optimized schedules. Piper achieves 2-3.5X higher MFU than state-of-the-art frameworks such as X-MoE, and a novel all-to-all algorithm delivers 1.2-9X bandwidth over vendor implementation.

Large language models (LLMs) have shown promise as interactive agents that solve tasks through extended sequences of environment interactions. While prior work has primarily focused on system-level optimizations or algorithmic improvements, the role of task horizon length in shaping training dynamics remains poorly understood. In this work, we present a systematic empirical study that examines horizon length through controlled task constructions. Specifically, we construct controlled tasks in which agents face identical decision rules and reasoning structures, but differ only in the length of action sequences required for successful completion. Our results reveal that increasing horizon length alone constitutes a training bottleneck, inducing severe training instability driven by exploration difficulties and credit assignment challenges. We demonstrate that horizon reduction is a key principle to address this limitation, stabilizing training and achieving better performance in long-horizon tasks. Moreover, we find that horizon reduction is related to stronger generalization across horizon lengths: models trained under reduced horizons generalize more effectively to longer-horizon variants at inference time, a phenomenon we refer to as horizon generalization.

We introduce a finite-size gradient-transport framework for real language-model training, based on five observables $(D,z,β,δ,v_{\mathrm{rel}})$ that separate cascade size, duration, absolute transport, and intensive transport efficiency. We analyze direct raw-gradient measurements from Pico-LM across four scales and 125 aligned steps, together with a five-scale Pythia companion dataset built from 153 aligned checkpoint-difference update fields. The same algebraic closure holds in both families, and both share a near-unity cascade-size backbone, but they occupy distinct transport regimes: Pico-LM shows positive duration scaling and negative intensive-efficiency scaling, whereas Pythia remains near the $D=1$ baseline with only weak positive efficiency scale dependence. Randomized-field controls give nearly matched null floors in the intensive and duration channels, indicating that the contrast reflects different real departures from a shared null skeleton rather than different null calibrations. The families also differ in stepwise power-law compressibility: Pico-LM retains clean duration and efficiency power laws, whereas Pythia preserves the size backbone but shows weaker one-slope compressibility in those channels. External performance associations are correspondingly channel-level, carried mainly by $v_{\mathrm{rel}}$ and normalized cascade duration, while $D(t)$ acts as a shared size backbone without a significant exponent-level performance association. These results support a reusable transport measurement framework without claiming a universal fixed point or a first-principles derivation of neural scaling laws.

Low-rank decomposition is a compelling approach for compressing large language models, but its effectiveness hinges on selecting which singular-vector bases to retain for a target task. Existing methods such as Basel adapt singular-value coefficients on downstream data and prune bases with small re-learned magnitudes, a heuristic that can be misaligned with task performance because it ignores the local geometry of the loss landscape. We present Basis Selection with Importance (BSI), a principled low-rank compression framework that ranks and prunes bases by directly estimating the expected loss increase incurred when each basis is removed. BSI derives a derivative-based importance score from a second-order Taylor expansion of the task loss with respect to singular values, combining first-order sensitivity and second-order curvature to quantify pruning impact. To make this criterion practical for LLMs, we develop an efficient Hessian-diagonal estimator by adapting the Hutchinson randomized-probing method to loss curvature with symmetric parameter perturbations. We provide a comprehensive theoretical analysis, including loss-increase bounds under basis pruning, explicit propagation of Hessian-diagonal estimation error into these bounds, variance characterization tied to the Hessian spectrum, high-probability sample-complexity guarantees for achieving a target estimation accuracy, and guidance on perturbation intensity. Extensive experiments on mathematical reasoning benchmarks demonstrate that BSI consistently outperforms state-of-the-art low-rank decomposition baselines, with especially strong improvements under deep compression.

Training large language models (LLMs) is highly memory-intensive, as training must store not only weights and optimizer states but also intermediate activations for backpropagation. While existing memory-efficient methods largely focus on gradients and optimizer states, activation compression is less well established due to the lack of LLM-tailored theory and guarantees. In this work, we develop a theoretical framework showing that activation compression is safe for linear operators when activation compression is unbiased, but problematic for nonlinear ones. We further derive gradient variance bound and establish convergence guarantees for applying activation compression to all linear operators under the standard $L$-smoothness assumption, showing that it does not change the convergence rate. Guided by the theory, we propose an activation-gradient co-compression method that reuses low-rank activation factors to compress linear-layer gradients without extra computation or additional gradient error. We conduct extensive experiments on Qwen and LLaMA models using a pretraining benchmark and multiple fine-tuning benchmarks to validate our theory and demonstrate competitive performance of our method in both accuracy and compression efficiency. We provide our code in the supplementary material for reproducibility.

Scaling laws enable the optimal selection of data amount and language model size, yet the impact of the data unit, the token, on this relationship remains underexplored. In this work, we systematically investigate how the information granularity of tokens, controlled by the compression rate (i.e., average bytes of text per token), affects scaling trends. We train 988 latent tokenized models (BLT) ranging from 50M to 7B parameters that enable setting the desired compression rate. This flexibility allows us to study the role of compression rate well beyond 4.57 bytes per token obtained with a popular BPE tokenizer. Our experiments reveal that in compute-optimal configurations, model parameter counts scale proportionally to data size measured in bytes, not in tokens as commonly perceived (Kaplan et al., 2020; Hoffmann et al., 2022). Furthermore, we discover that the optimal compression rate differs from the one obtained with BPE and decreases with compute. These findings generalize to both latent and subword tokenization, as well as to languages other than English, guiding language model developers on tokenization scheme selection for maximal compute efficiency.

In the era of big data, large-scale machine learning models have revolutionized various fields, driving significant advancements. However, large-scale model training demands high financial and computational resources, which are only affordable by a few technological giants and well-funded institutions. In this case, common users like mobile users, the real creators of valuable data, are often excluded from fully benefiting due to the barriers, while the current methods for accessing large-scale models either limit user ownership or lack sustainability. This growing gap highlights the urgent need for a collaborative model training approach, allowing common users to train and share models. However, existing collaborative model training paradigms, especially federated learning (FL), primarily focus on data privacy and group-based model aggregation. To this end, this paper intends to address this issue by proposing a novel training paradigm named decentralized relay learning (DeRelayL), a sustainable learning system where permissionless participants can contribute to model training in a relay-like manner and share the model. In detail, this paper presents the architecture and workflow of DeRelayL, designs incentive mechanisms to ensure sustainability, and conducts theoretical analysis and numerical simulations to demonstrate its effectiveness.

Code sandboxes have emerged as a critical infrastructure for advancing the coding capabilities of large language models, providing verifiable feedback for both RL training and evaluation. However, existing systems fail to provide accurate verification and efficiency under high-concurrency workloads. We present ScaleBox, a high-fidelity and scalable system designed to address these limitations in large-scale code training. ScaleBox introduces automated special-judge generation and management, fine-grained parallel execution across test cases with seamless multi-node coordination, and a configuration-driven evaluation suite for reproducible benchmarking. A series of experiments demonstrates that ScaleBox significantly enhances code verification accuracy and efficiency. Our further RLVR experiments show that ScaleBox substantially improves both performance on LiveCodeBench and training stability, significantly outperforming heuristic-matching baselines. By providing a reliable and high-throughput infrastructure, ScaleBox facilitates more effective research and development in large-scale code training.

Can a small group of volunteers shape how AI systems discuss animal welfare, just by editing Wikipedia? We show that they can. Wikipedia appears in nearly every major language model training dataset and is weighted more heavily than web-crawled text. The Pro-Animal Wikipedians (PAW), a group of advocates who add sourced animal welfare content to relevant articles, have made 125 edits across 115 pages. Using gradient-based data attribution (Bergson; MAGIC), we traced how these edits influence language model behavior. TrackStar retrieval attribution on Llama 3.1 8B found that PAW-edited sections made up 68 percent of the highest-attributed documents for animal welfare queries (p < 0.0001) but only 52 percent for unrelated queries about the same companies (p = 0.53): the model links PAW content specifically to animal welfare topics, not to the entities in general. MAGIC counterfactual influence estimation on Llama-3.2-1B, run across five random training-order seeds, gave the same picture even more sharply: in every seed, the top-10 most influential documents on animal welfare queries were all PAW edits (10 of 10, 5 of 5 seeds), while on general queries the same top-10 sat at chance (4 to 6 of 10). Mean PAW influence exceeded mean control influence on animal welfare queries with p < 0.0001 in every seed, an effect 6 to 30 times larger than on general queries. Leave-subset-out validation gave Spearman rho = 1.00 for all 10 runs. When we fine-tuned separate models on PAW content versus control content, each model performed better specifically on the type of text it was trained on: the PAW-trained model cut perplexity on animal welfare text from 12.4 to 8.4, while the control-trained model cut perplexity on control text from 16.1 to 11.4. A small, coordinated Wikipedia editing campaign therefore measurably shapes how language models handle the topics those edits address.

Large language models (LLMs) and multimodal large models (MLLMs) are increasingly used for transportation tasks such as regulation question answering, traffic management support, engineering review, and autonomous-driving scene reasoning. Yet transportation workflows are rule-intensive, computation-intensive, safety-critical, and inherently multimodal. Existing general benchmarks provide limited evidence of whether a model can apply regulations correctly, perform verifiable engineering calculations, or interpret traffic scenes reliably, while the small number of public transportation benchmarks remain narrow in scope and rarely support fine-grained diagnosis across text, images, and point-cloud data. To address this gap, we present TRIP-Evaluate, an open multimodal benchmark for large models in transportation. The benchmark organizes 837 items using a role-task-knowledge taxonomy that covers vehicle, traffic-management, traveler, and planning-and-design functions. Each item is annotated with capability, modality, and difficulty labels, enabling diagnosis from overall accuracy down to specific failure modes. The current release includes 596 text items, 198 image items, and 43 point-cloud items. TRIP-Evaluate also standardizes item construction, quality control, prompting, decoding, and scoring to improve cross-model comparability. Results on a diverse panel of models show that text-based performance is improving, but substantial weaknesses remain in multi-step engineering calculation, rule-constrained reasoning, multimodal scene understanding, and point-cloud understanding. Overall, TRIP-Evaluate provides a reproducible, diagnosable, and engineering-aligned evaluation baseline for model selection, regression testing, and safer deployment in transportation applications.

An effective way to scale up test-time compute of large language models is to sample multiple responses and then select the best one, as in Grok Heavy and Gemini Deep Think. Existing selection methods often rely on external reward models, which requires training a strong reward model and introduces additional computation overhead. As an alternative, previous approaches have explored intrinsic signals, such as confidence and entropy, but these signals are noisy with naive aggregation. In this work, we observe that high-entropy tokens tend to cluster into consecutive groups during inference, providing a more stable notion of model uncertainty than individual tokens. Together, these clusters reveal temporal patterns of model uncertainty throughout the inference process. Motivated by this observation, we propose to use the temporal structure of uncertainty as an intrinsic reward. To this end, we first formalize the basic unit of segment-level uncertainty as the High Entropy Phase (HEP), a variable-length segment that begins at a high-entropy token and ends when consecutive low-entropy tokens appear. We then define the Entropy Centroid, inspired by the concept of the center of mass in physics, as the weighted average position of all HEPs along the trajectory. Intuitively, a lower centroid indicates early exploration followed by confident generation, which we find often corresponds to higher response quality. Based on this insight, we propose the Lowest Centroid method, which selects the response with the lowest entropy centroid among multiple candidates. Experiments on mathematics, code generation, logical reasoning, and agentic tasks, across model scales ranging from 14B to 480B, show that Lowest Centroid consistently outperforms existing baselines and delivers stable gains as model size increases. Code is available at https://github.com/hkust-nlp/entropy-centroid.

Test-time scaling (TTS) has emerged as a powerful paradigm for improving the reasoning performance of large language models (LLMs) by investing additional compute at inference time. A central component of TTS is the \emph{verifier}, which selects or scores candidate solutions to guide the search process. While prior work has explored the benefit of verification, a fundamental question remains underexplored: \emph{what is the optimal granularity of verification under a given compute budget?} Coarse-grained outcome reward models (ORMs) and fine-grained process reward models (PRMs) represent two extremes, yet neither alone achieves compute-optimality across all regimes. In this paper, we establish a unified theoretical framework, called \textbf{GRACE} (\underline{G}ranularity-\underline{R}egulated \underline{A}daptive \underline{C}omputational \underline{E}fficiency), that characterizes the optimal verification granularity as an explicit function of problem difficulty, verifier accuracy, and compute budget. We prove that there exists a phase transition: fine-grained verification dominates when either the compute budget is large or the problem is hard, whereas coarse-grained verification is preferred in the low-budget, easy-problem regime. Our theory unifies Best-of-$N$, beam search, and step-level MCTS within a single Pareto-optimality framework, and motivates an adaptive granularity strategy that provably achieves the compute-performance Pareto frontier. Empirical results on MATH-500, GSM8K, and AIME benchmarks corroborate all four theoretical claims, with our adaptive strategy outperforming fixed-granularity baselines by up to 3.1% accuracy at matched compute.

Scaling test-time compute has emerged as a powerful mechanism for enhancing Large Language Model (LLM) performance. However, standard post-training paradigms, Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), optimize the likelihood of individual samples under a base policy, creating a misalignment with test time procedures that rely on aggregated or filtered outputs. In this work, we propose Compute Aligned Training, which aligns training objectives with test-time strategies. By conceptualizing inference strategies as operators on the base policy, we derive new loss functions that maximize performance when said strategies are applied. We instantiate such loss functions for SFT and RL across common test time strategies. Finally, we provide empirical evidence that this training method substantially improves test time scaling over standard training.

Training large neural networks with data-parallel stochastic gradient descent allocates N GPU replicas to compute effectively identical updates -- a practice that leaves the rich space of learning rate configurations entirely unexplored during training. We propose Hyperparameter-Divergent Ensemble Training (HDET), a method that repurposes these replicas for simultaneous learning rate exploration at negligible communication overhead. HDET operates in alternating phases: a fan-out stage in which replicas train independently under a structured, symmetric spread of learning rates, and a converge stage in which parameters are averaged across all replicas via AllReduce every T steps. Building on this ensemble substrate, we further propose an automatic learning rate (auto-LR) controller that treats the relative training loss across replicas as a performance signal, updating the shared base schedule toward higher-performing configurations via a momentum-based gradient-free meta-update. The combined method produces a self-adapting learning rate schedule that improves both optimization quality and generalization without additional hyperparameter sweeps or training budget. Crucially, the framework generalizes beyond learning rate: any scalar hyperparameter that does not alter model architecture -- such as dropout rate, attention scale temperature, or weight-decay coefficient -- can be explored across replicas using the same fan-out/converge protocol, with inter-replica loss differences serving as zero-order hypergradients that guide the search direction. HDET is implemented as a drop-in replacement for PyTorch's OneCycleLR scheduler, requiring no changes to model architecture, optimizer, or data pipeline.

Modern Deep Learning Recommendation Models (DLRMs) follow scaling laws with sequence length, driving the frontier toward ultra-long User Interaction History (UIH). However, the industry-standard "Fat Row" paradigm, which pre-materializes these sequences into every training example, creates a storage and I/O wall where data infrastructure usage exceeds GPU training capacity due to data redundancy that is amplified in multi-tenant environments where models with vastly different sequence length requirements share a union dataset. We present a \emph{versioned late materialization} paradigm that eliminates this redundancy by storing UIH once in a normalized, immutable tier and reconstructing sequences just-in-time during training via lightweight versioned pointers. The system ensures Online-to-Offline (O2O) consistency through a bifurcated protocol that prevents future leakage across both streaming and batch training, while a read-optimized immutable storage layer provides multi-dimensional projection pushdown for heterogeneous model tenants. Disaggregated data preprocessing with pipelined I/O prefetching and data-affinity optimizations masks the latency of training-time sequence reconstruction, keeping training throughput compute-bound by GPUs. Deployed on production DLRMs, the system reduces training data infrastructure resource usage while enabling aggressive sequence length scaling that delivers significant model quality gains, serving as the foundational data infrastructure for modern recommendation model architectures, including HSTU and ULTRA-HSTU.

Large Language Models (LLMs) have achieved strong performance across natural language and multimodal tasks, yet their practical deployment remains constrained by inference latency and kernel launch overhead, particularly in interactive, short-sequence settings. This paper presents a hybrid runtime framework that combines Just-In-Time (JIT) compilation with CUDA Graph execution to reduce launch overhead while preserving runtime flexibility during autoregressive decoding. The framework partitions transformer inference into static components executed via CUDA Graph replay and dynamic components handled through JIT-compiled kernels, enabling asynchronous graph capture and reuse across decoding steps. We evaluate the proposed approach on LLaMA-2 7B using single-GPU, batch-size-one inference across prompt lengths from 10 to 500 tokens. Experimental results show that the hybrid runtime reduces Time-to-First-Token (TTFT) by up to 66.0% and achieves lower P99 latency compared with TensorRT-LLM in this regime. These results indicate that hybrid JIT-CUDA Graph execution can effectively reduce inference latency and variance for short-sequence LLM workloads, making it a practical optimization strategy for latency-sensitive AI applications.

Quantum computing devices are recognized as powerful tools for solving NP-complete problems. However, the intricacy of their modeling presents notable barriers for non-specialists, while the tedious iteration of constraint weights and modeling methodologies also consumes substantial effort on the part of experts. To address these challenges, this study integrates a femtosecond laser-pumped Coherent Ising Machine (CIM) with an LLM-driven agentic system by leveraging the LangGraph and LangChain frameworks. Comprehensive investigations demonstrate that large language models (LLMs) can effectively perform such tasks in modeling as QUBO/Ising model calibration, constraint weight decision iteration and rapid validation of literature-reported schemes. Notably, all these tasks can be fully implemented based on domestic large models, combined with domestically developed CIM hardware, we truly achieve the practical empowerment of quantum CIM that fully relies on all-domestic agentic large models and hardware. This work successfully realizes robust technological integration, laying a solid foundation for subsequent research. Nevertheless, it also identifies the persisting challenges in the two cutting-edge fields of large models and quantum computing at the current stage. Encouragingly, we unexpectedly discover a promising new paradigm where accumulated knowledge from agent-assisted quantum computing iterations reciprocally enhances the agent's own problem-solving capability, thereby addressing these challenges.

Low-rank adaptation (LoRA) has emerged as the de facto standard for parameter-efficient fine-tuning (PEFT) of foundation models, enabling the adaptation of billion-parameter networks with minimal computational and memory overhead. Despite its empirical success and rapid proliferation of variants, it remains elusive which architectural choices, optimization techniques, and deployment constraints should guide practical method selection. This overview revisits LoRA through the lens of signal processing (SP), bridging modern adapter designs with classical low-rank modeling tools and inverse problems, as well as highlighting how SP principles can inform principled advances of fine-tuning approaches. Rather than providing a comprehensive enumeration and empirical comparisons of LoRA variants, emphasis is placed on the technical mechanisms underpinning these approaches to justify their effectiveness. These advances are categorized into three complementary axes: architectural design, efficient optimization, and pertinent applications. The first axis builds on singular value decomposition (SVD)-based factorization, rank-augmentation constructions, and cross-layer tensorization, while the second axis deals with initialization, alternating solvers, gauge-invariant optimization, and parameterization-aware methods. Beyond fine-tuning, emerging applications of LoRA are accounted across the entire lifecycle of large models, ranging from pre- and post-training to serving/deployment. Finally, open research directions are outlined at the confluence of SP and deep learning to catalyze a bidirectional frontier: classical SP tools provide a principled vocabulary for designing principled PEFT methods, while the unique challenges facing modern deep learning, especially the overwhelming scale and prohibitive overhead, also offer new research lines benefiting the SP community in return.

We measure how much one recurrence is worth to a looped (depth-recurrent) transformer, in equivalent unique parameters. From an iso-depth pretraining sweep across recurrence counts $r \in \{1, 2, 4, 8\}$ spanning ${\sim}50\times$ in training compute, we fit a joint scaling law $L = E + A\,(N_\text{once} + r^{\varphi} N_\text{rec})^{-α} + B\,D^{-β}$ and measure a recurrence-equivalence exponent $\varphi = 0.46$. Intuitively, $\varphi$ tells us whether looping a block $r$ times is equivalent in validation loss to $r$ unique blocks of a non-looped model (full equivalence, $\varphi{=}1$) or to a single block run repeatedly with no capacity gain ($\varphi{=}0$). Our $\varphi = 0.46$ sits in between, so replacing unique blocks with shared recurrences increases validation loss at matched training compute. For example, at $r{=}4$ a 410M looped model performs on par with a 580M non-looped model, but incurs the training cost of a 1B non-looped one. We demonstrate the utility of $\varphi$ as a diagnostic tool on two case studies: commonly used truncated backpropagation lowers $\varphi$ to $0.38$, indicating that the loop mechanism is poorly trained under truncation, even though validation loss decreases. Conversely, hyperconnections raise $\varphi$ to $0.65$, a genuine capacity gain. Our method separates true loop improvements from training-side gains, a distinction raw validation loss cannot make.

Language models are increasingly used in scientific discovery to generate hypotheses, propose candidate solutions, implement systems, and iteratively refine them. At the core of these trial-and-error loops lies evaluation: the process of obtaining feedback on candidate solutions via verifiers, simulators, or task-specific scoring functions. While prior work has highlighted the importance of evaluation, it has not explicitly formulated the problem of how evaluation-driven discovery loops can be scaled up in a principled and effective manner to push the boundaries of scientific discovery, a problem this paper seeks to address. We introduce Simple Test-time Evaluation-driven Scaling (SimpleTES), a general framework that strategically combines parallel exploration, feedback-driven refinement, and local selection, revealing substantial gains unlocked by scaling evaluation-driven discovery loops along the right dimensions. Across 21 scientific problems spanning six domains, SimpleTES discovers state-of-the-art solutions using gpt-oss models, consistently outperforming both frontier-model baselines and sophisticated optimization pipelines. Particularly, we sped up the widely used LASSO algorithm by over 2x, designed quantum circuit routing policies that reduce gate overhead by 24.5%, and discovered new Erdos minimum overlap constructions that surpass the best-known results. Beyond novel discoveries, SimpleTES produces trajectory-level histories that naturally supervise feedback-driven learning. When post-trained on successful trajectories, models not only improve efficiency on seen problems but also generalize to unseen problems, discovering solutions that base models fail to uncover. Together, our results establish effective evaluation-driven loop scaling as a central axis for advancing LLM-driven scientific discovery, and provide a simple yet practical framework for realizing these gains.

Mixture-of-Experts (MoE) has become the dominant architecture for scaling large language models: frontier models routinely decouple total parameters from per-token computation through sparse expert routing. Scaling laws show that under fixed active computation, model quality scales predictably with total parameters, and MoEs realize this by increasing expert count. However, training large MoEs is expensive, as memory requirements and inter-device communication both scale with total parameter count. We propose expert upcycling, a method for progressively expanding MoE capacity by increasing the number of experts during continued pre-training (CPT). Given a trained E-expert model, the upcycling operator constructs an mE-expert model through expert duplication and router extension while holding top-K routing fixed, preserving per-token inference cost. Duplication provides a warm initialization: the expanded model inherits the source checkpoint's learned representations, starting from a substantially lower loss than random initialization. Subsequent CPT then breaks the symmetry among duplicated experts to drive specialization. We formalize the upcycling operator and develop a theoretical framework decomposing the quality gap into a capacity term and an initialization term. We further introduce utility-based expert selection, which uses gradient-based importance scores to guide non-uniform duplication, more than tripling gap closure when CPT is limited. In our 7B-13B total parameter experiments, the upcycled model matches the fixed-size baseline on validation loss while saving 32% of GPU hours. Comprehensive ablations across model scales, activation ratios, MoE architectures, and training budgets yield a practical recipe for deploying expert upcycling, establishing it as a principled, compute-efficient alternative to training large MoE models from scratch.

Large language model optimization has historically bifurcated into isolated data-centric and model-centric paradigms: the former manipulates involved samples through selection, augmentation, or poisoning, while the latter tunes model weights via masking, quantization, or low-rank adaptation. This paper establishes a unified \emph{data-parameter correspondence} revealing these seemingly disparate operations as dual manifestations of the same geometric structure on the statistical manifold $\mathcal{M}$. Grounded in the Fisher-Rao metric $g_{ij}(θ)$ and Legendre duality between natural ($θ$) and expectation ($η$) parameters, we identify three fundamental correspondences spanning the model lifecycle: 1. Geometric correspondence: data pruning and parameter sparsification equivalently reduce manifold volume via dual coordinate constraints; 2. Low-rank correspondence: in-context learning (ICL) and LoRA adaptation explore identical subspaces on the Grassmannian $\mathcal{G}(r,d)$, with $k$-shot samples geometrically equivalent to rank-$r$ updates; 3. Security-privacy correspondence: adversarial attacks exhibit cooperative amplification between data poisoning and parameter backdoors, whereas protective mechanisms follow cascading attenuation where data compression multiplicatively enhances parameter privacy. Extending from training through post-training compression to inference, this framework provides mathematical formalization for cross-community methodology transfer, demonstrating that cooperative optimization integrating data and parameter modalities may outperform isolated approaches across efficiency, robustness, and privacy dimensions.

Recent works proposed test-time alignment methods that rely on a small aligned model as a proxy that guides the generation of a larger base (unaligned) model. The implicit reward approach skews the large model distribution, whereas the nudging approach defers the generation of the next token to the small aligned model when the large base one is unconfident about its outcome. In this work, we first show that both approaches can be reduced to sampling from similar graphical models, where they differ only in the definition of a rejection criterion (or distribution). Moreover, we argue that the confidence criterion is ill-motivated due to linguistic phenomena like ambiguous phrasing. We propose a novel rejection criterion based on a conservative confidence bet. Experimentally, our novel approach outperforms previous work on several datasets.

Parallel reasoning enhances Large Reasoning Models (LRMs) but incurs prohibitive costs due to futile paths caused by early errors. To mitigate this, path pruning at the prefix level is essential, yet existing research remains fragmented without a standardized framework. In this work, we propose the first systematic taxonomy of path pruning, categorizing methods by their signal source (internal vs. external) and learnability (learnable vs. non-learnable). This classification reveals the unexplored potential of learnable internal methods, motivating our proposal of STOP (Super TOken for Pruning). Extensive evaluations across LRMs ranging from 1.5B to 20B parameters demonstrate that STOP achieves superior effectiveness and efficiency compared to existing baselines. Furthermore, we rigorously validate the scalability of STOP under varying compute budgets - for instance, boosting GPT-OSS-20B accuracy on AIME25 from 84% to nearly 90% under fixed compute budgets. Finally, we distill our findings into formalized empirical guidelines to facilitate optimal real-world deployment. Code, data and models are available at https://bijiaxihh.github.io/STOP

Universal Machine Learning Interatomic Potentials (uMLIPs), pre-trained on massively diverse datasets encompassing inorganic materials and organic molecules across the entire periodic table, serve as foundational models for quantum-accurate physical simulations. However, uMLIP training requires second-order derivatives, which lack corresponding parallel training frameworks; moreover, scaling to the billion-parameter regime causes explosive growth in computation and communication overhead, making its training a tremendous challenge. We introduce MatRIS-MoE, a billion-parameter Mixture-of-Experts model built upon invariant architecture, and {Janus}, a pioneering high-dimensional distributed training framework for uMLIPs with hardware-aware optimizations. Deployed across two Exascale supercomputers, our code attains a peak performance of 1.2/1.0 EFLOPS (24%/{35.5%} of theoretical peak) in single precision at over 90% parallel efficiency, compressing the training of billion-parameter uMLIPs from weeks to hours. This work establishes a new high-water mark for AI-for-Science (AI4S) foundation models at Exascale and provides essential infrastructure for rapid scientific discovery.

Test-time scaling (TTS) improves large language models (LLMs) by allocating additional compute at inference time. In practice, TTS is often achieved through parallel scaling: generating multiple candidate responses and selecting the best via a Best-of-N (BoN) strategy. Its effectiveness therefore hinges on the scoring function. Learned scorers such as process reward models (PRMs) can be strong, but they are expensive to train and run. Lightweight confidence heuristics based on token log-probabilities are much cheaper, yet we find that they often perform substantially worse. To improve on lightweight confidence heuristics without incurring the full cost of stronger learned scorers, we introduce SCATR, a simple and efficient BoN ranking method that learns a lightweight scorer from a small calibration set using hidden representations from the base model. Across coding and mathematical reasoning benchmarks, SCATR improves over prior confidence-based baselines by up to 9%. Relative to LoRA fine-tuning on the same calibration data, it achieves comparable accuracy with up to 8000x fewer trainable parameters and much lower compute, reducing training and inference latency by up to 150x and 1000x, respectively. SCATR is also competitive with strong PRM baselines, and in several settings improves accuracy by up to 7.8% on math and 4.2% on coding while enabling up to 1000x faster inference. Overall, SCATR offers a strong accuracy-efficiency trade-off for scalable test-time selection.

We motivate the use of differentiable probabilistic programming techniques in order to account for the large model-space inherent to astrophysical $γ$-ray analyses. Targeting the longstanding Galactic Center $γ$-ray Excess (GCE) puzzle, we construct differentiable forward model and likelihood that make liberal use of GPU acceleration and vectorization in order to simultaneously account for a continuum of possible spatial morphologies consistent with the GCE emission in a fully probabilistic manner. Our setup allows for efficient inference over the large model space using variational methods. Beyond application to $γ$-ray data, a goal of this work is to showcase how differentiable probabilistic programming can be used as a tool to enable flexible analyses of astrophysical datasets.