Oussama Hidaoui, Omer Ebead, Ulrich Armel Mbou Sob +14cs.LG cs.AI
Generalising to unseen tasks remains a fundamental challenge in offline multi-agent reinforcement learning (MARL). In this work, we present a principled analysis of zero-shot task generalisation in the offline setting and conduct an extensive empirical investigation into the scaling behaviour governing task diversity, dataset size, and network capacity. To facilitate this study, we extend offline sequence modelling architectures to handle multi-task observation and action spaces alongside variable agent counts across tasks. Our primary finding is that scaling task diversity---rather than sheer dataset size is the dominant factor in achieving robust zero-shot transfer. Through large-scale experiments across four challenging environments (Connector, RWARE, SMAX, and LBF), we demonstrate that our multi-task approach achieves a mean improvement of 3.2x on held-out test tasks compared to single-task models and consistently outperforms strong behaviour cloning baselines. These results suggest that the development of generalisable MARL agents should prioritise the diversity of the training distribution with varying numbers of agents, providing a roadmap for scaling offline MARL effectively.
Scaling laws in modern deep learning describe how held-out loss improves as model capacity, training data, and compute increase, often following power-law trends. We investigate whether analogous scaling regularities arise in actuarial ratemaking, where data are tabular, heterogeneous, and noisy, and where classical models such as GLMs remain strong baselines. Using a real-world motor insurance portfolio, we train models from different families across increasing fractions of the training data and multiple random seeds, evaluating out-of-sample Poisson deviance, a likelihood-based loss for Poisson count predictions in which lower values indicate better held-out fit. We find that all model families improve with additional data, but scaling exponents differ substantially: TabM exhibits markedly stronger data scaling than purely supervised tabular Transformers and standard MLP baselines. Transformer variants show weak parameter scaling unless augmented with additional inductive biases (TabM-style adaptation or self-supervision). These results provide quantitative guidance on model selection by data regime and suggest that effective scaling on actuarial tabular tasks depends on architecture and loss function objective design, with simple increases in Transformer size providing limited gains.
Looped Transformers increase effective depth by iterating a shared block of layers, but most evaluations compare at fixed model size, conflating architectural advantage with extra FLOPs. We study looping on Mixture-of-Experts Transformers while closely matching per-token FLOPs, total non-embedding parameters, and KV cache. Through a series of ablations, we arrive at a recipe we call SMELT (Sparse MoE Transformer, middle layers Loop Twice), which loops the middle half of layers twice while matching the unlooped Baseline on all three budgets. We scale SMELT across four sizes up to 54B non-embedding parameters and fit a separate Chinchilla-style scaling law for each architecture. SMELT's loss drops faster with compute, saving 6.8--18.0\% of training FLOPs on the compute-optimal frontier. The advantage transfers to downstream benchmarks beyond what validation loss predicts, is largest on Code, and grows with sample length and the number of in-context examples. Mechanistic analysis shows that the second visit reduces the attention sink and redirects mass toward content-relevant tokens, an inductive bias that may underlie the observed performance gains. These results show that looping can improve Transformers even under budget matching, offering a practical recipe that turns depth reuse into measurable gains.
Reinforcement learning with verifiable rewards (RLVR) improves language-model reasoning, but how these gains relate to inference-time decoding and search remains unclear. Does RL create reasoning the base model lacks, or shift the rollout distribution toward trajectories it can already reach but rarely samples? We study this behaviorally with a Unified Decoding Framework (UDF), which expresses token-level sampling, beam-like search, tree search, and sequence-level resampling as executable policies over a shared budgeted operating space, scored post hoc with pass@$k$, self-consistency, best-of-$N$, and first-finish success. Using paired Base/RL checkpoints from SimpleRL-Zoo, we ask whether an RL default-policy curve can be approximated by a structured path of Base operating points. On Math500, AIME, GPQA, and IFEval, the pass@$k$ recovery path follows a Budgeted Operating-Point Transition Rule (BOPTR), $N_{\mathrm{Base}} \approx αN_{\mathrm{RL}}^β$, with benchmark-conditioned exponents. On Qwen2.5-7B, BOPTR gives the lowest transfer error among the non-oracle rules we test, 3.41 pp (95% CI [2.32, 5.53]); a three-seed replication gives 3.07 $\pm$ 0.39 pp. The rule extends to ten models across four families (3.28 to 4.87 pp on checkpoints added after fitting), to four benchmarks it was never fitted on (5.03 pp vs. 4.44 pp in fit), and holds without an RL checkpoint for the target model (4.19 pp) or without RL supervision of any kind (5.08 pp). These results support a qualified internalized-search reading: under the recipe we test, much of the measured RL gain corresponds to a change in sampling efficiency toward operating points the base model can already reach under search. We treat the scaling patterns as descriptive of this recipe and cohort, report where they break down, and use UDF and BOPTR as behavioral diagnostics rather than evidence of parameter-level equivalence.
Recent advances in large reasoning models (LRMs) have shown that reinforcement learning with verifiable rewards (RLVR) can substantially improve reasoning in mathematics and code, where outcomes can be checked automatically. Extending this progress to open-ended and agentic tasks remains difficult because reliable rewards are harder to obtain and direct human supervision cannot keep pace with the scale and complexity of model-generated experience. This paper studies how LRMs can continue to improve as human supervision gradually recedes from the learning loop. We examine two connected dimensions of this problem. The reward axis traces the development from per-instance human judgments to reusable verifiers and rewards that operate even without human feedback. The experience axis examines how learning can progress from human-curated tasks and environments toward self-generated curricula, constructed environments, and autonomous co-evolution. We connect these dimensions through a five-level ladder from L0 to L4 that identifies which parts of the learning process remain under continued human control. Our analysis further highlights the risks introduced by increasingly autonomous rewards and experience generation, including reward hacking, feedback drift, curriculum collapse, and environment errors. Consequently, we also provide the evaluation around three complementary objects: policy capability, feedback fidelity, and experience quality. This analysis provides a structured account of current approaches to scaling LRMs beyond human supervision and the open problems involved in developing self-sustaining learning systems toward superintelligence. Furthermore, we maintain a continuously updated \href{https://github.com/visitworld123/Awesome-Scaling-LRM-Beyond-Human-Supervision}{GitHub repository} to track the latest advances.
Vage Taamazyan, Zhuowen Shen, Stefan Hinterstoisser +6cs.CV
Stereo reconstruction is one of the last remaining Computer Vision tasks where all state-of-the-art methods employ a heavy architectural inductive bias. Even though it has been demonstrated that the task can be solved using general-purpose methods, it is widely believed that inductive biases in stereo are strictly necessary for both high-quality results and computational efficiency. We challenge this paradigm. In this paper, we demonstrate that both state-of-the-art accuracy and superior runtime efficiency are achievable with a model completely devoid of architectural inductive biases, relying instead on a simple, end-to-end Vision Transformer. By training on massive synthetic datasets, we show that pure data-driven learning can surpass explicitly engineered geometry. This work proves that explicit inductive biases are no longer a prerequisite for stereo matching, ultimately unlocking true scaling laws for continuous improvement in 3D reconstruction.
Video generation for autonomous driving cannot follow the web-scale route: driving data is expensive to collect, bound by privacy requirements, and cannot be scraped at will, so models must make the most of a fixed corpus. We present a systematic scaling-law study of video diffusion models trained from scratch on driving data: a family of models from 1M to 9B parameters, trained at different exposures on up to 5,500 hours of driving. Validation loss follows consistent power laws in both model size and training exposure, answering the questions that shape a training budget: whether compute is better spent on longer training or on a larger model, and whether more data is needed. Loss improves much faster with training exposure than with model size, making longer training the most effective way to improve a fixed model under limited compute. However, larger models continue to achieve lower asymptotic loss, so compute-optimal scaling still favors increasing model size when sufficient compute and data are available. Guided by these laws, we train a 9B-parameter model, to our knowledge the largest video diffusion model trained from scratch on driving data: it sets a new open-source state of the art for driving video generation, as measured on nuScenes. Our code and pretrained models are available at https://github.com/valeoai/VATIX. NATIX is separately releasing the underlying driving data in stages.
Niccolò Ajroldi, Diana Alexandra Onutu, Haider Al-Tahan +4cs.LG cs.AI
We study the scaling behavior of learning rate and batch size in pretraining dense large language models on English-prevalent corpora. Beyond scaling jointly optimal learning rates and batch sizes, we investigate their marginal evolution with model capacity and data scale and develop a model that captures these relationships. As we employ a Warmup-Stable-Decay learning rate schedule, we further investigate the gains from learning rate annealing over a broad range of hyperparameters settings, models and data budgets, and whether the optimal learning rate and batch size transfer between the stable and decay phases. Finally, we characterize the dependence of loss on model capacity and dataset size, evaluating recently proposed scaling forms that explicitly model their interaction. We find these approaches particularly effective at capturing both undertraining and overtraining regimes across our experiments. This study establishes a first baseline and scaling procedure for the development of future OpenEuroLLM models. We open-source the complete collection of pretraining runs used in this study.
Scaling laws are usually read as a capability story: lower language-modeling loss yields more useful models. We study a safety consequence of this mechanism in \emph{cross-session decomposition attacks}, where benign-looking subqueries are asked across independent interactions and later recomposed toward a forbidden objective. We formalize this setting as \emph{compositional safety risk} and prove a conditional risk-transfer bound: when the reference environment already contains dispersed evidence for a risky reconstruction, the gap between deployed composed risk and reference composed risk is controlled by the model's excess loss on allowed subqueries. Synthetic withholding experiments show that wider transformers assign lower loss to held-out instructions that never appear verbatim in training but are recoverable from injected supporting facts. A 600-intent pretrained-LLM evaluation shows that larger Qwen3 and Gemma3 family members can yield greater harmful-capability uplift under a fixed decomposition-composition pipeline. As a defense, IntentAlign-MiniLM, our 22M-parameter intent-aligned retriever, outperforms much larger embedding models on held-out intent retrieval and yields the best learned-retriever harmful recall across tested guardrails. Code is available in \href{https://github.com/liaodisen/Cross-Session-Decomposition-Attacks}{our GitHub repository}.
Serdar Gülbahar, Lukas Edman, Alexander Frasercs.CL cs.LG
Pre-training under limited data requires a different view of scaling than web-scale language modeling. With a fixed data budget but relatively abundant compute, increasing parameter count helps only up to an optimal scale; beyond that point, models overfit and generalization worsens. We study this behavior across 10M-100M word pre-training budgets, two corpora, and multiple downstream evaluations, and find that optimal size depends strongly on both the data budget and the downstream target. We argue that standard Transformers scale down poorly to this setting, because embeddings consume a large fraction of the parameter budget and per-token computation is tied to representational capacity. To address this coupling, we study recursive Transformers, reusing a shared block across depth to scale compute, together with factorized embeddings to reduce vocabulary-map parameters. We train three recursive models and find that they outperform standard Transformers at 10M and 100M words, while remaining competitive with BabyLM Challenge 2025 winners.
Nayeon Kim, Hojin Lee, Yunju Bak +2cs.LG cs.AI cs.CL
Mixture-of-Experts (MoE) architectures significantly expand model capacity without a proportional increase in computational cost. However, optimizing their hyperparameters---particularly the learning rate---at extreme scales of both model size and token budget via sweeping remains computationally prohibitive. In this paper, we propose a compute-efficient, two-step hyperparameter transfer framework that estimates optimal learning rates for training large MoE models by transferring them across scaling model widths, and subsequently extrapolating to trillion-token horizons. First, we formulate a Maximal Update Parameterization ($μ$P) adaptation for MoE architectures utilizing Multi-head Latent Attention (MLA) and the Muon optimizer, demonstrating that optimal learning rates transfer consistently across width-scaled models. Second, we extend this transferability along the token dimension by establishing a predictive scaling law. By applying linear regression to the optimal values derived from small proxy models on limited budgets, we successfully extrapolate the ideal learning rate to massive training horizons (e.g., 10 trillion tokens) with high fidelity ($R^2=0.95$). Consequently, this indicates that proxy training on small models is sufficient to determine the optimal learning rate for the extensive training of large-scale MoEs. We apply the proposed methodology to pretrain our foundation model (155B total, 17B active parameters) from scratch, and the stable training and evaluation results validate that optimal configurations for full-scale target models can be accurately predicted with minimal ablation costs.
Compute-optimal scaling laws guide the training of frontier language models yet remain largely unexplored for visual generation. We present a systematic scaling law study for text-to-image diffusion models using Abra, a controlled family of flow-matching transformers trained across three orders of magnitude worth of compute ($10^{19}$ to $10^{22}$ FLOPs), reaching significantly larger compute budgets than previous works. We demonstrate that diffusion models scale just as predictably as language models but require far more data to train optimally: compute optimality occurs at approximately $200$ image tokens per parameter, ten times the Chinchilla compute-optimal prescription for LLMs. We show that unlike language models, diffusion models are robust to overtraining and that practitioners should err on the side of more data rather than a larger model. Finally, we show that this predictability extends beyond training loss to generative quality metrics, optimal CFG settings, representation quality, and even the shape of the training curves, which collapse onto a universal form.
Children learn hundreds of words over the first years of their lives, in a process that begins slowly but quickly picks up speed. Prior models describe vocabulary growth as evidence accumulation over time. Here we show that the process is best characterized as accelerating accumulation: children learn more from each additional unit of linguistic experience than they did from the one before. In contrast to children, language models -- even those trained on child-directed speech -- do not accelerate. Instead, they show constant proportional returns on new data, consistent with scaling laws. Children learn using many orders of magnitude less training data than language models; their increasingly efficient use of their learning input is a candidate explanation.
Test-time compute can substantially improve Large Language Model (LLM) reasoning performance, yet how and when additional compute helps remains poorly understood. We study Divergent-Convergent Reasoning (DCR), a simple two-phase primitive consisting of an exploration phase that generates multiple candidate solutions followed by a convergent reconciliation phase. We present three core results. First, we show that even a single reconciliation step can reliably amplify correct minority reports: across datasets, DCR often recovers the correct answer when correct exploration outputs are in the minority, a regime where majority voting fails. Second, we introduce recursive DCR, an autoregressive reconciliation system that iteratively analyzes disagreements and allocates additional test-time compute. Recursive DCR achieves higher accuracy than fixed-compute baselines-reaching 93.3% on AIME 2024 and 92.0% on AIME 2025-while using roughly 27% less compute on average, demonstrating that attentive resource allocation is superior to uniform scaling. Third, we analyze disagreement among exploration outputs via a simple, training-free dispersion metric. Dispersion reveals a structured relationship between disagreement and test-time gains: in regimes where DCR is effective, higher disagreement among exploration outputs is associated with larger accuracy improvements from reconciliation. Together, these results show that disagreement, often viewed as noise, can be systematically exploited to improve test-time reasoning and reveal emerging scaling laws for agentic LLM systems.
As large language models scale, their training-token budgets must also increase to maintain an appropriate tokens-per-parameter ratio (\(\mathrm{TPP}\)). However, high-quality domain data is much harder to scale than general web data. As model size and the training-token budget increase, its fraction in the training mixture tends to decrease. Repeating the available high-quality data provides an effective way to counteract this dilution, but excessive repetition may lead to overfitting. We study this trade-off under practical LLM scaling, where the training-token budget grows proportionally with model size. For a fixed domain, we first find that, surprisingly at a fixed \(\mathrm{TPP}\), the optimal repetition count mildly increases with model size. Across different domains, we find that the optimal repetition count is strongly negatively correlated with the final validation loss of a domain: domains with lower loss can generally benefit from more repetitions. In contrast, the amount of unique domain data is only weakly related to the optimal repetition count. These findings suggest that repetition counts tuned on smaller proxy models with the same \(\mathrm{TPP}\) can provide a practical estimate for larger models.
Artificial Intelligence (AI) safety systems combine character shaping (e.g., Reinforcement Learning from Human Feedback [RLHF], Constitutional AI), which modifies behavioral distributions at training time, with rule enforcement (e.g., output filters, safety classifiers), which blocks harmful outputs at inference time, yet little formal analysis exists on how their optimal balance should change as deployment scales increase. We introduce a stylized comparative-statics model that parameterizes safety design as a resource allocation alpha in [0,1] between these two approaches, incorporating scale-dependent filter degradation, common-mode failures, and character fragility -- the risk that shaped behavior degrades or collapses under novel conditions. Under a multiplicative Pareto damage model, we derive closed-form expected harm and supplement it with tail-risk (CVaR) analysis via Monte Carlo simulation. Across three scenarios (optimistic, moderate, pessimistic), the optimal alpha* is interior or at the rules-only boundary and shifts weakly toward character shaping as deployment scale T grows, from negligible (Delta alpha* = +0.01) to pronounced (Delta alpha* = +0.21) depending on scenario. The dominant parameter is the baseline character fragility rate p^(0)_frag, which shifts alpha* by 0.50 across its range -- far exceeding the effect of tail severity, filter quality, or common-mode failure probability. CVaR and expected-harm optima converge at large T. These results suggest that safety architecture decisions depend less on deployment scale per se than on the reliability of character shaping under distributional shift.
Liu Ziyin, Yizhou Xu, Tomaso Poggio +1cs.LG cond-mat.dis-nn cond-mat.stat-mech
Neural networks trained by gradient descent on a smooth cost function can nevertheless learn in steps: the cost holds on long plateaus and then drops abruptly. Meanwhile, training losses instead follow smooth power laws. Variants of both behaviors occur in architectures with very different microscopic structures, which is the signature of a few relevant collective variables. We show that a symmetry fixes what those variables are: a network layer is a sum over interchangeable units, so relabeling the units leaves it unchanged; given smoothness and the condition that a unit's gradient vanish at the origin, symmetry then enforces a universal leading form for the expansion about the near-zero weights present at the start of training, the quadratic $\Tr[WW^{\top}A(x)]$, in which every architectural detail is confined to a single ``structure matrix" $A(x)$ that we compute for each architecture. Perceptrons, attention layers, mixtures of experts, and convolutions become one model at different $A$. Its training dynamics then close on the ``order parameter" $M=WW^{\top}$ and, whenever the data matrices share an eigenbasis, reduce to a Lotka--Volterra equation whose modes switch on one after another. The smaller the initial weights, the further apart the switch-on times, and the plateaus appear as a singular limit of a smooth flow; when many modes are unresolved the same events merge into a power law in training time whose exponent the theory predicts. We confirm both numerically across training methods and architectures.
Nicholas Lourie, Kyunghyun Cho, Karen Ullrich +1cs.LG
Scaling laws promised cost-effective experiments; six years later, they have yet to fully deliver. Instead, researchers have found them unreliable at small scales (starting at 4M parameters) and concluded that sizable models cannot be avoided. We show this is not the case: the confounding factor is hyperparameters. Small models are highly sensitive, but hyperparameter sensitivity fades with scale. This small-scale sensitivity makes scaling laws easy to miss because they only emerge on the fully tuned frontier, and reaching that frontier requires an extensive search far beyond what most ever run. By ablating the basic scaling law recipe, we show well-tuned hyperparameters matter more than any other ingredient. Further, we reveal why those hyperparameters become easier to find: as scale increases, the hyperparameter loss surface becomes lower dimensional. Nevertheless while scaling laws exist in small models, extrapolation hits statistical limitations. A holistic approach is required. Synthesizing our insights with the recent literature, we develop a new methodology for model-centric research and demonstrate it on a question that once took the field years to settle: where to place normalization layers in the transformer architecture. From small-scale experiments, we recover the large scale result: pre-normalization works better as models grow in size. With the right tools and a better understanding, small-scale experiments can deliver on scaling laws' long-awaited promise.
Ellen Su, Andres Potapczynski, Shikai Qiu +2cs.LG cs.CL
Modern systems are increasingly expected to transfer across tasks not specified during training. What data facilitates generalization in these new, unanticipated settings? One hypothesis is that data with more structural information could contain shared circuits and subprograms that could be recycled in a wider array of downstream settings. Epiplexity, a recently proposed measure of the structural information a compute-bounded learner can extract from data, provides a mechanism to reason about this relationship. In this paper, we show how to operationalize epiplexity as an online training signal for data selection and synthetic data generation. For selection, we fit scaling laws to the training loss curves of natural data domains to predict the expected epiplexity gain as a function of training tokens, and use this signal to adaptively determine the sampling weights over domains during training. For synthetic data generation, we define a generator's reward as the change in learner epiplexity over a buffer of previously generated data and use REINFORCE policy gradients to guide the generator toward an epiplexity-maximizing distribution. In both cases, higher epiplexity predicts improved downstream performance on zero-shot and fine-tuning based tasks, supporting the hypothesis that data rich in structural information yield representations that transfer across domains.
In large-scale pretraining, the algorithm, architecture, and systems decisions are conventionally made in disconnected stages. A scaling law stage selects an architecture and training recipe, optimizing loss under compute constraints, and a separate systems stage then optimizes the implementation for hardware efficiency. In this work, we develop MOSAIC, which formulates model architecture and systems co-design as an optimization problem. MOSAIC couples a predictive scaling law with a calibrated performance model that estimates Model FLOPs Utilization (MFU), communication cost, memory footprint, and the best parallel layout. We instantiate the framework for sparse Mixture-of-Experts (MoE) language models, where expert count, routing sparsity, and other MoE layer dimensions affect both the loss and systems efficiency. We fit a scaling law on sparse MoE models trained on text data, whose scaling dimensions include the sparsity factor, which is the fraction of model parameters inactive per token in a forward pass. The scaling law sweeps in our work span active parameters from $104$ million to $2.7$ billion and total model sizes reaching $79$ billion parameters. We show that, within the calibrated sparsity range, an efficiency-agnostic model-FLOPs budget admits no interior optimal sparsity. The fitted loss decreases monotonically with sparser models and the compute optimum lies at the upper boundary of the data support. An optimal sparsity in MoE models instead emerges under the cluster's systems constraints, as captured by MOSAIC. Our results argue for a shift towards unified architecture and systems co-design for frontier language model training.
Mathurin Videau, Badr Youbi-Idrissi, David Lopez-Paz +1cs.CL
Neural scaling laws are foundational for language model development, yet standard formulations systematically under- and overestimate loss at data-scarce and overtraining extremes. This failure originates in the underlying assumption that model size and training data impact the loss independently. To address this, we introduce the Skaling law, a generalized functional form that couples model capacity and data through a single interaction exponent. This simple extension reduces the Mean Absolute Percentage Error (MAPE) by 1.5-3x across both interpolation and extrapolation regimes. When paired with a sparse grid strategy restricted to low-compute regimes, the Skaling law achieves accurate full-grid extrapolation using approximately 10x less compute than uniform sweeps. By enabling reliable performance prediction from small-scale experiments, the Skaling law provides a more robust and resource-efficient framework for allocating compute budgets in next-generation model training.
Aggregate scaling curves suggest that Video LLMs improve smoothly or saturate as visual budgets grow. We show that this view can conceal large, opposing changes at the item level. We represent each frozen model--item pair by its response trajectory under controlled visual budgets and derive matched-grid measures of configuration complementarity, harmful transitions, and text overwrite. Across five open Video LLMs from three architecture families, four multiple-choice benchmark splits, open-ended QA and summarization, and fixed-history dialogue generation, no single budget serves all items. On the four-model matched MCQA grid, item-level oracle headroom spans $8.8$--$18.9$ accuracy points and $12.5$--$25.5\%$ of items are correct at a lower budget but wrong at a higher one. Task-appropriate continuous metrics show the same complementarity beyond multiple choice: Token-F1 oracle gaps are $2.7$--$3.7$ score points on MLVU generation and $3.8$--$4.8$ points on AVSD current-turn generation, even when mean quality improves with budget. The effect persists across frame count, spatial resolution, sampling policy, temporal--spatial allocation, and independently executed raw-video and cached pipelines, with per-item rates and membership tracking protocol choices. A controlled sampling intervention recovers $29.0\%$ of terminal regressions, and a structured frame audit identifies several recurring evidence pathways. We release per-item trajectories, protocol provenance, derived annotations, and reproducible analysis code as an auditing artifact. A confidence cascade matches fixed-$128f$ accuracy while reducing average shared frame cost by $31.7\%$, illustrating one operational use of the response matrix.
Large language models fine-tuned on human behavioural data have emerged as general-purpose cognitive proxies, but the scale this requires, and whether these models process task structure or exploit statistical shortcuts, remain open questions. We train fourteen models from 135M to 14B parameters across four architecture families on Psych-101, a dataset of 10.7 million trial-level choices from 160 experiments. For in-distribution simulations, scale barely matters. The models fall within a narrow band, as though against a ceiling, and 0.6B to 1B parameters suffice to match a 70B baseline on held-out participants. Out-of-distribution, that band opens into a markedly steeper scaling gradient, with larger models clearly advantaged in generalisation to novel task structure. To determine what information these models use, we run two diagnostics. We progressively strip four prompt channels -- task instructions, experimental stimuli, outcome feedback, and choice history -- across 27 experiments, and permute trial order. Masking the content of stimuli and feedback destroys 75.7% of learned information and pushes models below chance, demonstrating that choice history alone does not account for performance. Permutation reveals invariance on tasks with independent trials but sensitivity where trial order is determined by prior responses. Small cognitively fine-tuned models therefore show promise as noise ceiling estimators for psychological experiments, though their scope remains bounded by the paradigms seen in training.
Diffusion language models (dLLMs) offer an alternative to autoregressive (AR) language modeling, yet the scaling behavior of Mixture-of-Experts (MoE) dLLMs remains poorly understood. We systematically characterize how optimization hyperparameters, compute allocation, and architecture scale for MoE dLLMs, identifying quantitative differences from scaling trends previously reported for AR models. Specifically, for optimization, the optimal nominal batch size grows faster, while the optimal learning rate decays more rapidly with compute. For model--data allocation, IsoFLOP analysis reveals a slight data-side tilt: the optimal token budget grows faster than activated model-side computation. For MoE architecture, larger scales increasingly favor larger expert pools at fixed activated capacity, while moderate expert granularity remains consistently effective and the preferred fraction of activated capacity assigned to shared experts remains stable across scales. Guided by these findings, we train LLaDA MoE v2, a 30B-A3B dLLM, from scratch on 23.5T tokens. With approximately 65\% as many pretraining tokens as Qwen3, LLaDA MoE v2 approaches Qwen3 on several knowledge, reasoning, and coding benchmarks. After supervised fine-tuning alone, it outperforms SDAR Chat on seven of eight reasoning and coding benchmarks and remains close to Qwen3 on several tasks. These results establish practical scaling laws and design principles for MoE dLLMs.
The discovery of scaling laws has highlighted the extraordinary potential of AI systems with a striking empirical pattern: as AI systems scale, their capabilities tend to improve predictably. Yet, in real-world applications, AI rarely operates in isolation; instead, it often works alongside humans, raising the question of whether these gains persist in human-AI collaboration. In this work, we develop an analytical model to examine when the empirical scaling benefits of AI translate into improved human-AI joint system performance. We demonstrate that the performance of a human-AI system can scale positively as the AI scales up-provided that humans have an accurate perception of the AI's capabilities. Human misperception, however, can fundamentally alter this relationship: i) when humans over-perceive the AI's capabilities, a scaling paradox may arise, in which greater AI scale reduces overall system performance and amplifies firm-level profit losses, and (ii) when humans under-perceive the AI's capabilities, performance still improves with scale but at a substantially slower rate. We further show that firms can actively manage these distortions through operational policies such as cost internalization and perception alignment, whose effectiveness depends on the economics of AI deployment and the direction of human misperception. These findings suggest that organizations may benefit more from managing the human-AI interface than from simply investing in larger, more expensive AI systems. More broadly, our results suggest that AI scaling should be viewed not only as a technological challenge, but also as a behavioral and operational one, and caution against the view that larger AI systems will automatically lead to better operational outcomes. Whether AI scaling creates value ultimately depends on how increased AI capabilities shape human beliefs and collaborative efforts.
We study empirical scaling properties for text conditioning in visual generation. Such properties have rarely been measured because diffusion loss does not scale with the number of tokens in natural-language prompts. Surprisingly, we find that the converged diffusion loss scales with the amount of structured language in the prompt. To quantify structured language, we adapt two complementary measures: a white-box likelihood metric (GPG) and a black-box attribute metric (ED). Across controlled training runs, the converged diffusion loss decreases approximately linearly with GPG and follows a power law with ED. Guided by these scaling properties, we improve \emph{diffusability} by constructing structured prompts with semantic and geometric annotations derived from images, and improve \emph{promptability} by training a prompter through supervised fine-tuning, cold-start, and verifier-gated on-policy distillation. The resulting system outperforms all evaluated open-weight models on nearly every compositional, reasoning, and world-knowledge benchmark, while matching or surpassing the strongest closed-weight models on most evaluations.
Visual generation increasingly requires high-resolution images, long videos, and multimodal context, making the quadratic cost of full attention prohibitive. We introduce Chimera, a hybrid visual diffusion backbone with a principled scaling recipe. Chimera processes text, image, and video tokens in one raster-ordered stream without positional embeddings. It combines Kimi Delta Attention (KDA) for long-context state tracking with O(N) complexity, interleaved Multi-head Latent Attention (MLA) for direct global interaction, and modality-aware short convolutions for local spatiotemporal context. Sparse Mixture-of-Experts (MoE) layers expand capacity while controlling activated compute. To scale this heterogeneous architecture, we introduce HeteroP, a module-wise scheme that transfers hyperparameters across width and depth according to each tensor's functional fan-in and model depth. HeteroP yields a consistently tuned family used to fit Chinchilla-style compute-optimal laws for activated model size, training-token count, and image-video data ratio. Guided by these laws, we train an 11B-parameter Chimera with 2B activated parameters. Experiments show three results. First, measured by pretraining diffusion loss, the dense backbone is 1.7x as compute-efficient as a matched full-attention Wan-2.1 2B baseline, while the complete system reaches 7.3x. Second, without length-specific fine-tuning, Chimera extrapolates zero-shot from 5-second training clips to 30-second videos, with only 6.5% FID degradation in the last five seconds. Third, the fitted laws show that compute-optimal image pretraining divides compute nearly evenly between activated model size and training-token count, whereas video pretraining modestly favors model size at higher budgets. These results establish a foundation for designing and scaling efficient long-context diffusion architectures.
Decoder-only language models entangle long-term memory and reasoning in a single parameter set, making it difficult to scale memory capacity independently. Memory Decoder introduces a parametric long-term memory module but only studies it at a relatively small scale. In this work, we present Memory Decoder at Scale, scaling memory models up to 6.9B parameters and pretraining them on 300B tokens. At this data scale, the combined cost of indexing and search makes a standard Faiss pipeline infeasible. We address this bottleneck with a distributed pipeline for Faiss indexing and retrieval, together with sparse, batch-wise loading of kNN distributions. Across model scales, we find that allocating more parameters to memory yields a better parameter-performance tradeoff than scaling the base model alone. On 17 benchmarks, pairing a 6.9B general memory with Pythia-410M raises its average score from 29.86 to 37.34, surpassing Pythia-12B (37.24) with 39% fewer total parameters. For Qwen3 Base models ranging from 0.6B to 14B, 1.7B domain memories improve the average score across the three domains by more than 9 points at every scale. Overall, our results demonstrate that independently scaling pretrained memory offers a more parameter efficient path to improving language model performance.
Jiaxiang Li, Zhiqi Bu, Shiyun Xucs.LG cs.AI math.OC
Modern deep learning typically keeps the batch size static throughout training, thus overlooking the joint effect of learning rate and batch size on the training dynamics. In this paper, we study the deep learning dynamics through the lens of convex optimization and derive a joint characterization of loss in terms of both schedules, applicable to general optimizers and model architectures. This characterization yields a closed-form optimal batch size schedule for any prescribed learning rate schedule, and further leads to joint scaling laws that consistently outperform static batch size baselines, highlighting the significance of dynamic batch size schedule in large language model training.
Tian Qin, Kimia Hamidieh, David Alvarez-Meliscs.LG cs.AI cs.PF
Classical compute-optimal scaling laws assume an unbounded supply of fresh pretraining data, yet pretraining is increasingly entering a regime in which compute grows faster than the availability of high-quality data. We propose Compute-Data (CD) scaling laws, a unified framework that bridges compute-optimal scaling, where data scales freely with compute, and data-optimal scaling, where the corpus is fixed while compute can grow without bound. CD scaling extends classical scaling laws by introducing a token-effectiveness function, $η$, which quantifies the value of a derived token-produced, for example, through multi-epoch repetition or paraphrasing-relative to a fresh token, ranging from a perfect substitute to having no value. We fit $η$ for two data-expansion strategies, multi-epoch repetition and paraphrasing, across model sizes from 14M to 600M parameters using the Dolma-3 corpus. We find that token effectiveness is far from constant: it depends jointly on model size, the tokens-per-parameter ratio, and the amount of derived data, and it saturates as the corpus is expanded. The functional form of $η$ implies diminishing returns when substituting compute for data as either model size or data availability increases. It also partitions training into three operational regimes---compute-bound, data-bound, and model-bound---and shows that classical compute-optimal allocation is suboptimal across most practically relevant settings.