Microsimulation models used by ministries of finance and central banks rely on parametric processes for lifetime earnings that capture only first and second moments of the conditional distribution and miss long-range nonlinear structure. We propose SAGA, a decoder-only transformer for irregular tabular panel sequences, paired with a split conformal calibration wrapper that delivers individual-level prediction intervals with finite-sample marginal coverage guarantees. Trained on the longitudinal Swedish LISA register over 1990 to 2022, comprising 2,143,817 individuals and 61,284,903 person-years, the model forecasts annual labor earnings at horizons of one to thirty years and aggregates them by Monte Carlo into present-discounted lifetime earnings distributions. Against the canonical Guvenen, Karahan, Ozkan, and Song parametric process and tabular and recurrent baselines, SAGA reduces continuous ranked probability score by 31.9 percent at the ten-year horizon and mean absolute error by 37.7 percent at the twenty-year horizon. Conformal intervals achieve nominal coverage to within 0.4 percentage points marginally and within 2.4 percentage points on the worst-case demographic subgroup. The reconstructed lifetime earnings Gini coefficient is 0.327 against the partially observed truth of 0.341 and the GKOS estimate of 0.378. Model weights, calibration tables, and a synthetic equivalent dataset are released for replication outside the protected SCB MONA environment.

Perplexity -- a function measuring a model's overall level of "surprise" when encountering a particular output -- has gained significant traction in recent years, both as a loss function and as a simple-to-compute metric of model quality. Prior studies have pointed out several limitations of perplexity, often from an empirical manner. Here we leverage recent results on Transformer continuity to show in a rigorous manner how perplexity may be an unsuitable metric for model selection. Specifically, we prove that, if there is any sequence that a compact decoder-only Transformer model predicts accurately and confidently -- a necessary pre-requisite for strong generalisation -- it must imply existence of another sequence with very low perplexity, but not predicted correctly by that same model. Further, by analytically studying iso-perplexity plots, we find that perplexity will not always select for the more accurate model -- rather, any increase in model confidence must be accompanied by a commensurate rise in accuracy for the new model to be selected.

Length generalization, the ability to generalize from small training context sizes to larger ones, is a critical challenge in the development of Transformer-based language models. Positional encoding (PE) has been identified as a major factor influencing length generalization, but the exact impact of different PE schemes on extrapolation in downstream tasks remains unclear. In this paper, we conduct a systematic empirical study comparing the length generalization performance of decoder-only Transformers with five different position encoding approaches including Absolute Position Embedding (APE), T5's Relative PE, ALiBi, and Rotary, in addition to Transformers without positional encoding (NoPE). Our evaluation encompasses a battery of reasoning and mathematical tasks. Our findings reveal that the most commonly used positional encoding methods, such as ALiBi, Rotary, and APE, are not well suited for length generalization in downstream tasks. More importantly, NoPE outperforms other explicit positional encoding methods while requiring no additional computation. We theoretically demonstrate that NoPE can represent both absolute and relative PEs, but when trained with SGD, it mostly resembles T5's relative PE attention patterns. Finally, we find that scratchpad is not always helpful to solve length generalization and its format highly impacts the model's performance. Overall, our work suggests that explicit position embeddings are not essential for decoder-only Transformers to generalize well to longer sequences.

We investigate the generalization properties of dense text embeddings when the embedding backbone is a large language model (LLM) versus when it is a non-LLM encoder, and we study the extent to which spherical linear interpolation (SLERP) model-merging mitigates over-specialization introduced by task-specific adaptation (e.g., LoRA). To make the comparison concrete and domain-agnostic, we design a controlled suite of experiments in which models embed short numerical sequences and are evaluated on their ability to cluster and classify those sequences according to well-defined number-theoretic properties. Our experimental protocol compares four families of models: (1) non-LLM encoders trained from scratch or fine-tuned for embeddings, (2) LLM-based encoders adapted with parameter-efficient methods (LoRA), (3) LLM-based encoders with LoRA followed by model souping merging into the base weights, and (4) the same LoRA-adapted LLMs merged using SLERP across checkpoints or stages. We evaluate representational quality with clustering indices (Silhouette and Davies Bouldin). We additionally analyze the use of kmeans labels to see if the embeddings encode any other information besides the one we are testing for. Empirically, we find that LLM-based backbones produce embeddings that better capture higher-order, compositional numeric patterns, but are prone to adapter dominance that degrades balanced generalization; SLERP merging consistently recovers base-model structure while retaining most task gains, yielding superior tradeoffs in clustering separability, and robustness compared to model souping or models that were not merged.

Large language models display in-context learning as an emergent effect of scale, but they rely on static weights during inference. In contrast, biological systems continually adapt via synaptic plasticity. We investigate whether explicit, biologically inspired plasticity can endow Transformers with faster in-sequence adaptation. To this end, we augment decoder-only Transformers with fast-weight modules updated either by (i) a neuromodulated Hebbian rule or (ii) the gradient-based plasticity mechanism of Duan et al. (2023). Across copying, regression, and few-shot classification tasks (CIF AR-FS, Omniglot), Hebbian plasticity consistently achieves lower loss and stronger few-shot generalization, while gradient-based updates perform best on long-horizon credit assignment. When associations are short and linearly separable, static weights suffice, defining a clear boundary condition for when plasticity helps. Analysis of learned modulatory signals reveals that gradient-based rules maintain large, persistent updates, whereas Hebbian plasticity is sharply gated around salient events. Together, these results show that explicit plasticity complements attention by enabling rapid, task-specific adaptation, and clarify when different plasticity mechanisms are most effective.

Work performed while at Google DeepMind.

We introduce Pixels2Play-0.1 (P2P0.1), a foundation model that learns to play a wide range of 3D video games with recognizable human-like behavior. Motivated by emerging consumer and developer use cases - AI teammates, controllable NPCs, personalized live-streamers, assistive testers - we argue that an agent must rely on the same pixel stream available to players and generalize to new titles with minimal game-specific engineering. P2P0.1 is trained end-to-end with behavior cloning: labeled demonstrations collected from instrumented human game-play are complemented by unlabeled public videos, to which we impute actions via an inverse-dynamics model. A decoder-only transformer with auto-regressive action output handles the large action space while remaining latency-friendly on a single consumer GPU. We report qualitative results showing competent play across simple Roblox and classic MS-DOS titles, ablations on unlabeled data, and outline the scaling and evaluation steps required to reach expert-level, text-conditioned control.

In this paper, we present Synergy, a language model that bridges different levels of abstraction in an end-to-end fashion through a learned routing mechanism. Focusing on low-level linguistic abstraction, we trained our model as a byte-level language model. Our model spontaneously learns to tokenize bytes, producing fewer concept tokens than Byte-level Byte Pair Encoder (BBPE) tokenizers while keeping comparable performance. By comparing with Llama3, we observed an advantage of Synergy under the same model scale and training dataset size. Further studies show that the middle part (the higher abstraction part) of our model performs better when positional encodings are removed, suggesting the emergence of position-independent concepts. These findings demonstrate the feasibility of tokenizer-free architectures, paving the way for more robust and flexible pipelines.

Real-world processes often generate data that are a mix of categorical and numeric values that are recorded at irregular and informative intervals. Discrete token-based approaches are limited in numeric representation capacity while methods like neural ordinary differential equations are not well suited for categorical data or informative sampling and require augmentation to handle certain classes of trajectories. Here, we present multivariateGPT, a single architecture for modeling sequences of mixed categorical (including tokenized text) and numeric data. This is accomplished with an autoregressive sequence decomposition, embedding scheme, and loss function that extend the next token prediction task to likelihood estimation of the joint distribution of next token class and value. We demonstrate how this approach can efficiently learn to generalize patterns in simple physical systems and model complex time series including electrocardiograms and multivariate electronic health record data. This work extends the utility of transformer based models to additional classes of data.