Puro-2B: Poor Lab's Qwen2-1.5B Trained on RTX 5090 within $5090

Language model pretraining has become almost synonymous with prohibitive cost, placing it out of reach for much of the academic and open-source communities. Although strong open-source efforts already exist, including open-weight models and open-source training recipes, a cost-efficient, hardware-accessible, and open-source pretraining recipe has long been missing. Even at a small scale, training Llama-3.2-3B costs over \$1.5M, and reproducing SmolLM3-3B needs over \$700K. In this report, we present an open pretraining recipe designed to lower this barrier. Using this recipe, we train a collection of Puro-2B models from scratch on up to 1.4 trillion tokens with FP8 precision on consumer-grade RTX 5090 GPUs. The models in the collection differ in token budgets and selected recipe variants. Our best model is trained at a compute cost of less than \$6.9K and approaches Qwen2.5-1.5B performance under our evaluation protocol. This cost efficiency is enabled by a combination of approaches, including hardware selection, low-precision training, hyperball optimization, curriculum model averaging, and the data recipe. Beyond the recipe itself, we provide two additional results. First, across the Puro-2B collection, we derive a Puro Cost Scaling Law that relates training cost to average model performance; the fitted law suggests that about \$4.4K, less than \$5,090, is sufficient to reach the performance of Qwen2-1.5B. Second, as an end-to-end case study, we examine how pretraining data curricula shape downstream performance after post-training. Such controlled studies are enabled by having access to the full pretraining pipeline rather than model weights alone. We release the full training recipe for Puro-2B, including data, code, and model weights under Apache 2.0 at https://huggingface.co/collections/thu-pacman/puro-2b.

Effective Learning Rate Governs Loss Dynamics in Language Model Pretraining

We uncover ELR collapse in language model pretraining: learning rate (LR) and parameter norm govern loss dynamics primarily through their ratio, the effective learning rate (ELR). When ELR is matched across runs, their loss trajectories collapse throughout training despite substantially different LRs and parameter norms. Across optimizers, architectures, datasets, and model scales, mean collapse errors are typically a few x 10^-3, below the seed-to-seed variation measured in a representative configuration. Systematic ablations identify normalization design and the timescale of LR-norm variation as key determinants of collapse precision. Controlled interventions further show that weight decay and Hyperball shape loss dynamics primarily through the ELR schedules they induce. Replacing LR with ELR enables a fitted functional scaling law (FSL) to transfer across norm-control methods. The resulting ELR-based FSL also explains delayed acceleration, a recurring effect of norm control. Together, these results establish ELR as a common coordinate linking LR scheduling, norm control, and loss dynamics.

Towards a Densing Law for User Representation Learning at Billion-Scale Capacity

User representation learning in real-world industrial scenarios is commonly scaled by increasing user amount, behavioral sequence length and model size. However, existing methods face two challenges: (i) Bottleneck for raw data scaling at billion-scale capacity, as performance exhibit diminishing performance gains with larger-scale raw text user behavioral input, which can be mitigated by tokenization. (ii) Lack of quantitative analysis of how tokenization configurations should scale with data size. In this report, we propose User Behavioral Densing Law for characterizing the quantitative relationship between data scale and the minimum sufficient tokenization capacity. Firstly, we conduct a pilot study on raw & tokenized scaling comparison on billion-scale Alipay dataset, revealing the raw data scaling bottleneck and the sustained gains enabled by tokenization. To derive the scaling pattern governing the minimum sufficient tokenization configuration at different data scales, theoretical analysis and systematic experiments are employed to summarize the quantitative scaling pattern. We find an approximately linear relationship between the logarithms of minimum sufficient tokenization capacity and input data size measured by tokens, and the scaling slope varies systematically with the tokenization method and data source, reflecting differences in representation-space redundancy and intra-source uniqueness. Guided by the proposed law, we further develop ALGN, an adaptive variable-length tokenization method that improves capacity allocation. Extensive experiments across diverse data sources, tokenization methods, and downstream tasks demonstrate the generalizability and reliability of the User Behavioral Densing Law, providing practical guidance for tokenization configuration selection in large-scale user representation learning. Moreover, ALGN outperforms existing baselines.

Teach and Grow: An Agent-Centered Architecture for General Robot Learning

End-to-end vision-language-action (VLA) and world-action models offer an elegant route to general-purpose robotics, but their reliability is bounded by validated physical coverage. When an unfamiliar object, sensor, embodiment, or contact falls outside that coverage and no validated fallback exists, correcting the failure requires new robot data, a policy update, and regression testing. This recurring burden is the retraining tax. Unlike text, embodied data must often be created by operating machines. We present Teach-and-Grow Learning (TGL), an agent-centered architecture for general robot learning. In its general form, a multimodal agent turns a few successful demonstrations into reusable Skill Blocks: closed-loop behaviors for meaningful subgoals. In a new scene, the agent grounds and composes these blocks, selects learned or geometric tools, observes the physical outcome, and revises the route when execution departs from intent. A Skill Library stores executable behavior, while structured Experience Memory carries forward success, failure, and repair. New tasks are acquired without task-specific policy retraining. Our LIBERO evaluation attains state-of-the-art performance; controlled studies expose skill induction, persistent reuse, and agent-directed adaptation. Finally, we propose the Teach-and-Grow scaling-law hypothesis: if X denotes effective reusable experience, future-task error and teaching demand should approach irreducible floors as power laws in X. The architecture therefore treats deployment as a period of continued learning, in which one task can make the next easier.

Scaling an Autoregressive Transformer for Single-Cell Generation

We study a self-supervised generation task for single-cell gene expression vectors: given a set of vectors from a cell type, we aim to generate additional gene expression vectors of that cell type. For this task we characterize both the biological fidelity of the generated gene expression vectors and the scaling behavior of the pretraining loss. The model is a causal transformer paired with a learned quantized VAE tokenizer, trained with a cross-entropy loss. To evaluate the model, we condition it on held-out gene expression vectors of a cell type and generate vectors of gene expression, comparing the resulting distribution over gene expression vectors to the ground truth distribution of that cell type. We study the scaling properties of the proposed architecture by varying the number of trained parameters and the amount of training data. To our knowledge, we find the first jointly-fit two-exponent scaling law and compute-optimal frontier for a single-cell foundation model. Finally, we discuss how this pretrained model could be finetuned for perturbation response prediction.

A scaling law of contextual persistence in human language

Human language exhibits lawful structure at the level of words (frequency, vocabulary growth) and word pairs (co-occurrence across distance). Here we show that the arrangement of words in sequence -- a central determinant of meaning -- obeys a comparable law. Using large language models as probabilistic probes, we measured the reduction in target perplexity conferred by prior context at distance d beyond that of the same words scrambled; this difference, the contextual persistence function P(d), isolates the influence of arrangement. Across ten corpora spanning six language families and written and spoken modalities, P(d) decayed approximately as 1/d ($P(d) \propto d^{-α}$, mean $α= 1.04$; median $r^2 = 0.96$). The effect vanished in scrambled and synthetic controls, replicated across independent probes, and did not appear in genomic or protein sequences under domain-native models. An exponent near 1 distributes contextual influence approximately uniformly across logarithmic timescales. The results establish a scaling law of contextual persistence in human language.

DomainPilot: Domain-Level Loss-Guided Two-Stage Data Mixture Optimization for Efficient Language Model Fine-Tuning

The training efficacy of large language models (LLMs) is fundamentally constrained by the quality and composition of training data. Existing dynamic data scheduling methods face critical limitations in industrial-scale pretraining and supervised fine-tuning (SFT): data selection incurs prohibitive O(N) costs on terabyte-scale corpora, mixture optimization schemes introduce severe I/O bottlenecks or require training auxiliary reference models, and sample-level reweighting strategies rely on loss signals that conflate noise, difficulty, and novelty. We present DomainPilot, a domain-level loss-guided two-stage data mixture optimization framework. DomainPilot introduces token-level domain loss monitoring to capture per-domain learning dynamics during training without halting the data pipeline. Building on these signals, we propose a Scaling Law guided coarse optimization stage that fits domain-specific convergence curves and derives a principled prior for mixture adjustment. A subsequent Mixing Law guided fine optimization stage refines the mixture by modeling cross-domain interaction effects through controlled sweep experiments. The entire mechanism is realized via a patch-based architecture that injects domain-aware loss computation into existing training frameworks (e.g., MindSpeed/Megatron-LM) with only ~30 lines of framework-specific adapter code. We validate DomainPilot on the Qwen3-1.7B model during SFT. Compared to the original data mixture, our optimized mixture achieves improvements of +2% on MMLU-Redux, +1.8% on AIME24, +3.8% on LiveCodeBench v5, and +3.6% on BFCL v3, without increasing total data volume or training cost. These results demonstrate that domain-level training signals provide an effective, lightweight alternative to expensive data selection or auxiliary model training for mixture optimization.

MESH: Scaling Up Retrieval with Heterogeneous Content Unification

Optimizing large-scale retrieval hinges on the ability to efficiently surface candidates across diverse content tiers. However, to capture segments such as fresh and long-tail content, modern systems typically resort to a fragmented "zoo" of specialized retrieval models. This operational complexity is attributed to a fundamental challenge in heterogeneous retrieval systems, the Scaling Bias of Heterogeneity, where model capacity gains do not apply equally across diverse content tiers. To bridge this gap, we propose MESH as a unified retrieval scaling framework that mitigates this bias through a modularized architecture integrated with gated bias correction. By partitioning the feature space into independent domains, MESH enforces a structural inductive bias that reduces interference between sparse-item signals and high-frequency engagement features. This protected gradient path leads to improved scaling behavior for sparse content, empirically validated by a 14 times improvement in the power-law scaling exponent for fresh items. In online evaluations on Pinterest's Related Pins platform, a billion scale item-to-item recommendation system, these improvements translate into a +5.5% lift in fresh-item repins, alongside with 55% improvement in funnel efficiency and +0.46% improvement in user retention. Finally, our asynchronous serving strategy ensures production viability by delivering a 2.87 times improvement in system throughput. Our findings suggest MESH as a promising paradigm for consolidating fragmented retrieval infrastructures into more scalable and ecosystem-aware backbones.

Data Predictability Shapes Weibull Weight-Scale Growth in Transformer Training

A trained transformer's weight magnitudes can be summarized by a two-parameter Weibull distribution whose shape $k \approx 1.2$ is stable across layers and models, so the scale $λ$ carries most training-induced movement. What corpus property sets how much $λ$ grows? Using the bigram conditional entropy $D = H(\text{next} \mid \text{prev})$, a training-free statistic computed before training, we find across controlled corruption families a learning-rate-conditioned law, $λ^2 - λ_0^2 = C_0(η) + C_1(η)(H_r - D)^{0.59}$, where $H_r$ is a matched-budget shuffle baseline. The convex exponent is inherited from an independently measured data-side saturation relation rather than fitted directly to the growth curve. After removing the two per-$η$ coefficients, 23 runs spanning an order of magnitude in learning rate collapse onto $(H_r - D)^{0.59}$ with unit slope ($R^2 = 0.941$; direct per-$η$ fits are weaker, $R^2 \approx 0.82$). Because $D$ is computed before training, the law is a forward predictor: an end-to-end self-validation recovers held-out within-family weight growth with 5.7% relative error. The readout holds at model and per-layer resolutions and across two tested architectures, with the functional form preserved and only the coefficients changing. It also marks its boundary: cross-corpus prediction over-predicts code, implicating redundancy as a second axis of a broader $Φ(D,R,A,H)$ data-to-weight framework.

Data-driven Machine Learning Cannot Reach Symbolic-level Logical Reasoning -- The Limit of the Scaling Law

By promoting vectors to spheres and enabling explicit model construction, neural networks can perform symbolic-level syllogistic reasoning without training data. We identify two fundamental limitations that prevent conventional data-driven machine learning systems from achieving this capability: training data generated by the combination table cannot distinguish all 24 valid syllogism types, and end-to-end premise-to-conclusion mapping creates contradictory targets within neural components. Experiments with two representative conventional systems, GPT-5 using linguistic inputs and Euler Net using visual inputs, support this analysis. ChatGPT GPT-5 may reach 100% accuracy in syllogistic reasoning, but with hallucinations. Because the learning process terminates upon reaching 100% accuracy, the system cannot progress beyond empirical accuracy to symbolic level reasoning. Random test data reduced Euler Net's accuracy to 56%. Repeatedly expanding the training set increased its accuracy to 97%, with perfect performance on 8 syllogism types. However, because unintended inputs cannot be exhaustively covered, even 100% test accuracy does not imply symbolic-level reasoning. Since syllogistic reasoning underpins logical reasoning and human rationality, these results suggest that increasing data and training time alone cannot ensure symbolic level logical reasoning.

The Energy Consumption of Transformer Fine-Tuning: A Roofline-Inspired Scaling Model

Transformer-based models underpin modern natural language processing but incur rapidly growing computational and energy costs. As training scales in both model size and parallelism, accurately predicting energy consumption has become critical for sustainable and cost-aware system design. We present a framework for modeling the energy consumption of Transformer training on multiple GPUs. Using controlled architectural sweeps of BERT models, we relate measured energy to lightweight proxies for compute, memory traffic, and hardware efficiency. Inspired by roofline models, our approach incorporates a speedup-based hardware-efficiency factor that captures the effects of tensor parallelism and fully sharded data parallelism. We derive a scaling law model that accurately predicts training energy across heterogeneous configurations.

Beyond Prompt Engineering: A Systematic Analysis of Prompt Lexical Sensitivity and Its Impacts on Quality

Large Language Models (LLMs) exhibit extreme sensitivity to surface-level prompt variations, in which minor lexical changes can trigger disproportionate performance fluctuations. Moving beyond black-box optimization and coarse-grained templates, we present the first large-scale, n-gram token-level mechanistic analysis of prompt stability, leveraging a dataset of 132,000 prompt variants. Our investigation reveals a fundamental Scaling Law of Prompt Performance Stability: higher average task performance is strongly associated with lower variance and greater robustness across prompt perturbation. We identify two core linguistic drivers underlying this robustness: (1) Domain-Specific Terminology, which tightly anchors semantic boundaries, and (2) Explicit Action Directives, which formalize reasoning trajectories. Together, these elements constrain the model's interpretative space, effectively ``locking in'' more deterministic generation behavior. Building on these insights, we introduce an automated Prompt-Refining Agent that systematically restructures input queries by injecting domain anchoring and operational constraints. Empirical evaluation shows that our approach reduces performance variance by 40.7% in code generation task, while preserving or improving mean performance. These findings provide a statistically grounded and mechanistically interpretable framework for achieving robust prompt engineering.

Streaming Communication in Multi-Agent Reasoning

Multi-agent reasoning systems adopt a "generate-then-transfer" paradigm that forces end-to-end latency to scale linearly with pipeline depth. We introduce StreamMA, a multi-agent reasoning system that streams each reasoning step to downstream agents as soon as it is generated, pipelining adjacent agents and thus reducing latency. Surprisingly, this pipelining also improves effectiveness: because multi-step reasoning quality is non-uniform and early steps are more reliable than later ones, working with these reliable early steps instead of the full chain prevents error-prone late steps from misleading downstream agents. We formalize both advantages with the first closed-form joint analysis of stream, serial, and single protocols, deriving the effectiveness ordering, speedup upper bound, and cost ratio. Across eight reasoning benchmarks spanning mathematics, science, and code, two frontier LLMs (Claude Opus 4.6 and GPT-5.4), and three topologies (Chain, Tree, Graph), StreamMA outperforms both baselines (avg. +7.3 pp, max +22.4 pp on HMMT 2026; Claude Opus 4.6-high). Beyond these contributions, we discover a "step-level scaling law": increasing per-agent steps consistently improves both effectiveness and efficiency, a new scaling dimension orthogonal to and composable with agent-count scaling.

Spectra as Language: Large Language Models for Scalable Stellar Parameter and Abundance Inference

Stellar spectra encode key information on the physical properties and chemical compositions of stars. Accurate stellar parameter determination is essential for addressing major questions such as galaxy and stellar evolution. Large-scale spectroscopic surveys have accumulated unprecedented spectral data. Traditional feature extraction or model-fitting approaches struggle with high-dimensional, massive datasets, limited generalization, and computational inefficiency. Recent advances in large language models demonstrate strong generalization and feature-learning in tasks like natural language processing, DNA/RNA sequence analysis, and protein/chemical parsing. Stellar spectra are continuous sequential signals, enabling the transfer of language models to stellar spectroscopy. Here, we propose a two-stage large language model framework for stellar parameter inference, achieving accurate estimation of effective temperature, surface gravity, metallicity, and abundances of ~20 chemical elements. Scaling-law analyses show systematic performance improvements with increasing data, providing a scalable framework for forthcoming large-scale surveys.