This paper investigates the scaling properties of autoregressive next-pixel prediction, a simple, end-to-end yet under-explored framework for unified vision models. Starting with images at resolutions of 32x32, we train a family of Transformers using IsoFlops profiles across compute budgets up to 7e19 FLOPs and evaluate three distinct target metrics: next-pixel prediction objective, ImageNet classification accuracy, and generation quality measured by Fr'echet Distance. First, optimal scaling strategy is critically task-dependent. At a fixed 32x32 resolution alone, the optimal scaling properties for image classification and image generation diverge, where generation optimal setup requires the data size grow three to five times faster than for the classification optimal setup. Second, as image resolution increases, the optimal scaling strategy indicates that the model size must grow much faster than data size. Surprisingly, by projecting our findings, we discover that the primary bottleneck is compute rather than the amount of training data. As compute continues to grow four to five times annually, we forecast the feasibility of pixel-by-pixel modeling of images within the next five years.

Large Language Models (LLMs) exhibit strong informal mathematical reasoning but struggle to generate mechanically verifiable proofs in formal languages like Lean. We present LEAP, an agentic framework that enables general-purpose foundation models to achieve state-of-the-art performance on automated formal theorem proving. LEAP leverages foundation model capabilities, such as informal reasoning, instruction following, and iterative self-refinement. By decomposing complex problems into smaller units, the system bridges formal proof construction with informal blueprints through continuous interaction with the Lean compiler. To provide a rigorous evaluation beyond increasingly saturated benchmarks, we introduce Lean-IMO-Bench, a benchmark of IMO-style problems formalized in Lean, with short statements yet highly non-routine and multi-step proofs across a wide range of difficulty levels. Empirically, on the latest 2025 Putnam Competition, an annual mathematics competition for undergraduate students in North America, LEAP solves all 12 problems, matching recent breakthroughs by frontier formal mathematical models. On Lean-IMO-Bench, LEAP boosts the one-shot formal solve rate of general-purpose LLMs from below 10

We present a case study in semi-autonomous mathematics discovery, using Gemini to systematically evaluate 700 conjectures labeled 'Open' in Bloom's Erdős Problems database. We employ a hybrid methodology: AI-driven natural language verification to narrow the search space, followed by human expert evaluation to gauge correctness and novelty. We address 13 problems that were marked 'Open' in the database: 5 through seemingly novel autonomous solutions, and 8 through identification of previous solutions in the existing literature. Our findings suggest that the 'Open' status of the problems was through obscurity rather than difficulty. We also identify and discuss issues arising in applying AI to math conjectures at scale, highlighting the difficulty of literature identification and the risk of ”subconscious plagiarism” by AI. We reflect on the takeaways from AI-assisted efforts on the Erdős Problems.

We present an extension and comprehensive real-world validation of Co-Scientist, a Gemini-based multi-agent system designed to accelerate end-to-end scientific research across hypothesis generation, experimentation, and manuscript generation. Moving beyond in silico hypothesis generation, this specialized configuration transitions Co-Scientist into an execution-grounded research partner advancing closed-loop scientific workflows across materials science, biology, and computer science. In materials science, Co-Scientist interfaced with a semi-automated chemical vapor deposition reactor to design a safe precursor route for MXenes; experimental execution produced a lamellar 2D material sharing key structural similarities with the Ti3C2Tx MXene lattice, although further experiments are needed to confirm the atomic structure. Leveraging Gemini 3 Deep Think for rapid, lab-in-the-loop execution, it also tailored growth recipes to laboratory constraints in minutes, enabling single-attempt growth of monolayer MoS2, MoSe2, and WS2 semiconductors. In biology, Co-Scientist predicted emergent swarming phenotypes of engineered E. coli across inducer (IPTG) gradients from sparse imaging data, quantitatively matching unpublished wet-lab morphological measurements. In computer science, Co-Scientist autonomously discovered an inference-time scaling architecture that outperformed six frontier models on HealthBench (Hard and Professional) while reducing potential clinical harm under blinded physician evaluation. Finally, a double-blind study of end-to-end generated papers with 30 domain experts across 450 reviews demonstrates that Co-Scientist's reliability modules reduce hallucination and plagiarism while improving research safety. Together, these results demonstrate progress toward closed-loop multi-agent scientific AI systems capable of accelerating real-world scientific discovery.

In medical education, physicians convert academic knowledge into clinical expertise through residency: years of training across thousands of encounters, with diverse sources of feedback and progressively greater autonomy. Much of clinical reasoning relies on the patient encounter, a dialogue in which a clinician elicits history, refines diagnostic hypotheses, and decides management under uncertainty. While large language models (LLMs) excel on static medical benchmarks, methods to optimize the full sequence of clinical decisions remain underdeveloped. We present ResidencyRL, a reinforcement learning (RL) method for training clinical artificial intelligence (AI) agents through simulated multi-turn clinical encounters (up to 60 dialogue turns and 8 tool calls per trajectory). ResidencyRL pairs the policy agent with LLM simulators capable of complex, adversarial behaviors, training against a structured reward aligned to diagnostic accuracy, management quality, communication, documentation, and safety. On held-out evaluations, the ResidencyRL agent improves diagnostic accuracy by 7.0

We report the performance of Aletheia (Feng et al., 2026b), a mathematics research agent powered by Gemini 3 Deep Think, on the inaugural FirstProof challenge. Within the allowed timeframe of the challenge, Aletheia autonomously solved 6 problems (2, 5, 7, 8, 9, 10) out of 10 according to majority expert assessments; we note that experts were not unanimous on Problem 8 (only). For full transparency, we explain our interpretation of FirstProof and disclose details about our experiments as well as our evaluation. Raw prompts and outputs are available at https://github.com/google-deepmind/superhuman/tree/main/aletheia.

Computers calculate transcendental functions by approximating them through the composition of a few limited-precision instructions. For example, an exponential can be calculated with a Taylor series. These approximation methods were developed over the centuries by mathematicians, who emphasized the attainability of arbitrary precision. Computers, however, operate on few limited precision types, such as the popular float32. In this study, we show that when aiming for limited precision, existing approximation methods can be outperformed by programs automatically discovered from scratch by a simple evolutionary algorithm. In particular, over real numbers, our method can approximate the exponential function reaching orders of magnitude more precision for a given number of operations when compared to previous approaches. More practically, over float32 numbers and constrained to less than 1 ULP of error, the same method attains a speedup over baselines by generating code that triggers better XLA/LLVM compilation paths. In other words, in both cases, evolution searched a vast space of possible programs, without knowledge of mathematics, to discover previously unknown optimized approximations to high precision, for the first time. We also give evidence that these results extend beyond the exponential. The ubiquity of transcendental functions suggests that our method has the potential to reduce the cost of scientific computing applications.

Recent advances in foundational models have yielded reasoning systems capable of achieving a gold-medal standard at the International Mathematical Olympiad. The transition from competition-level problem-solving to professional research, however, requires navigating vast literature and constructing long-horizon proofs. In this work, we introduce Aletheia, a math research agent that iteratively generates, verifies, and revises solutions end-to-end in natural language. Specifically, Aletheia is powered by an advanced version of Gemini Deep Think for challenging reasoning problems, a novel inference-time scaling law that extends beyond Olympiad-level problems, and intensive tool use to navigate the complexities of mathematical research. We demonstrate the capability of Aletheia from Olympiad problems to PhD-level exercises and most notably, through several distinct milestones in AI-assisted mathematics research: (a) a research paper (Feng26) generated by AI without any human intervention in calculating certain structure constants in arithmetic geometry called eigenweights; (b) a research paper (LeeSeo26) demonstrating human-AI collaboration in proving bounds on systems of interacting particles called independent sets; and (c) an extensive semi-autonomous evaluation (Feng et al., 2026a) of 700 open problems on Bloom's Erdos Conjectures database, including autonomous solutions to four open questions. In order to help the public better understand the developments pertaining to AI and mathematics, we suggest quantifying standard levels of autonomy and novelty of AI-assisted results, as well as propose a novel concept of human-AI interaction cards for transparency. We conclude with reflections on human-AI collaboration in mathematics and share all prompts as well as model outputs at https://github.com/google-deepmind/superhuman/tree/main/aletheia.

Finding the right north-star metrics is highly critical for advancing mathematical reasoning capabilities of foundation models, especially given that existing evaluations are either too easy or only focusing on getting correct short answers. To address these issues, we present IMO-Bench, a suite of advanced reasoning benchmarks that specifically targets the level of the International Mathematical Olympiad (IMO), the most prestigious venue for young mathematicians. IMOAnswerBench first tests models on 400 diverse Olympiad problems with verifiable short answers. IMO-ProofBench is the next-level evaluation for proof-writing capabilities, which includes both basic and advanced IMO problems as well as detailed grading guidelines to facilitate automatic grading. These benchmarks played a crucial role in our historic achievement of the gold-level performance at IMO 2025 with Gemini Deep Think (Luong and Lockhart, 2025). Our model achieved 80.0% on IMO-AnswerBench and 65.7% on the advanced IMO-ProofBench, surpassing the best non-Gemini models by large margins of 6.9% and 42.4% respectively. We also showed that autograders built with Gemini reasoning correlate well with human evaluations and construct IMO-GradingBench, with 1000 human gradings on proofs, to enable further progress in automatic evaluation of long-form answers. We hope that IMO-Bench will help the community towards advancing robust mathematical reasoning and release it at https://github.com/google-deepmind/superhuman/imobench.

Despite their success in many domains, large language models (LLMs) remain under-studied in scenarios requiring optimal decision-making under uncertainty. This is crucial as many real-world applications, ranging from personalized recommendations to healthcare interventions, demand that LLMs not only predict but also actively learn to make optimal decisions through exploration. In this work, we measure LLMs' (in)ability to make optimal decisions in bandits, a state-less reinforcement learning setting relevant to many applications. We develop a comprehensive suite of environments, including both context-free and contextual bandits with varying task difficulties, to benchmark LLMs' performance. Motivated by the existence of optimal exploration algorithms, we propose efficient ways to integrate this algorithmic knowledge into LLMs: by providing explicit algorithm-guided support during inference; and through algorithm distillation via in-context demonstrations and fine-tuning, using synthetic data generated from these algorithms. Impressively, these techniques allow us to achieve superior exploration performance with smaller models, surpassing larger models on various tasks. We conducted an extensive ablation study to shed light on various factors, such as task difficulty and data representation, that influence the efficiency of LLM exploration. Additionally, we conduct a rigorous analysis of the LLM's exploration efficiency using the concept of regret, linking its ability to explore to the model size and underlying algorithm.

Current benchmarks for large language model (LLM) reasoning predominantly focus on mathematical and coding abilities, leaving a gap in evaluating broader reasoning proficiencies. One particular exception is the BIG-Bench dataset, which has served as a crucial benchmark for evaluating the general reasoning capabilities of LLMs, thanks to its diverse set of challenging tasks that allowed for a comprehensive assessment of general reasoning across various skills within a unified framework. However, recent advances in LLMs have led to saturation on BIG-Bench, and its harder version BIG-Bench Hard (BBH). State-of-the-art models achieve near-perfect scores on many tasks in BBH, thus diminishing its utility. To address this limitation, we introduce BIG-Bench Extra Hard (BBEH), a new benchmark designed to push the boundaries of LLM reasoning evaluation. BBEH replaces each task in BBH with a novel task that probes a similar reasoning capability but exhibits significantly increased difficulty. We evaluate various general-purpose and reasoning-specialized models on BBEH and observe an accuracy of 23.9% for the best general-purpose model and 54.2% for the best reasoning-specialized model, indicating substantial room for improvement and highlighting the ongoing challenge of achieving robust general reasoning in LLMs. We release BBEH publicly at: https://github.com/google-deepmind/bbeh.

Existing reinforcement learning (RL) methods for large language models (LLMs) rely on static prompt sets, where prompts are curated a priori, and sampled in a fixed schedule for training, regardless of their usefulness to the RL process. We design , the first method that allows LLMs to prioritize and adaptively create useful prompts during RL training by reward signals. In principle, (Evolving via A symmetric Self-Play) casts language model training as a game between: (1) a creator, who samples and generates training prompts, and (2) a solver, who generates responses to the prompts. is simple, suits both offline and online RL for LLMs, and sets a new state-of-the-art on challenging benchmarks without extra human prompts: it improves gemma-2-9b-it’s win-rate on Arena-Hard from 51.6% to 60.1% by DPO and 52.6% to 62.4% by RLOO, surpassing claude-3-opus and nearing gemini-1.5-pro, both are orders of magnitude larger. Further ablation studies show can induce meaningful learning curriculum, and effectively scale RL for LLMs beyond static human prompts.

We present AlphaGeometry2 (AG2), a significantly improved version of AlphaGeometry introduced in (Trinh et al., 2024), which has now surpassed an average gold medalist in solving Olympiad geometry problems. To achieve this, we first extend the original AlphaGeometry language to tackle problems involving movements of objects, and problems containing linear equations of angles, ratios, and distances. This, together with support for non-constructive problems, has markedly improved the coverage rate of the AlphaGeometry language on International Math Olympiads (IMO) 2000-2024 geometry problems from 66% to 88%. The search process of AG2 has also been greatly improved through the use of Gemini architecture for better language modeling, and a novel knowledge-sharing mechanism that enables effective communication between search trees. Together with further enhancements to the symbolic engine and synthetic data generation, we have significantly boosted the overall solving rate of AG to 84% on all geometry problems over the last 25 years, compared to 54% previously. AG2 was also part of the system that achieved silver-medal standard at IMO 2024 https://dpmd.ai/imo-silver. Finally, we report progress towards using AG2 as a part of a fully automated system that reliably solves geometry problems from natural language input. Code: https://github.com/google-deepmind/alphageometry2.

Supervised fine-tuning (SFT) and reinforcement learning (RL) are widely used post-training techniques for foundation models. However, their roles in enhancing model generalization capabilities remain unclear. This paper studies the difference between SFT and RL on generalization and memorization, focusing on text-based rule variants and visual variants. We introduce GeneralPoints, an arithmetic reasoning card game, and adopt V-IRL, a real-world navigation environment, to assess how models trained with SFT and RL generalize to unseen variants in both textual and visual domains. We show that RL, especially when trained with an outcome-based reward, generalizes across both rule-based textual and visual variants. SFT, in contrast, tends to memorize training data and struggles to generalize out-of-distribution scenarios. Further analysis reveals that RL improves the model's underlying visual recognition capabilities, contributing to its enhanced generalization in the visual domain. Despite RL's superior generalization, we show that SFT remains essential for effective RL training; SFT stabilizes the model's output format, enabling subsequent RL to achieve its performance gains. These findings demonstrates the capability of RL for acquiring generalizable knowledge in complex, multi-modal tasks.

We present Step-Back Prompting, a simple prompting technique that enables LLMs to do abstractions to derive high-level concepts and first principles from instances containing specific details. Using the concepts and principles to guide the reasoning steps, LLMs significantly improve their abilities in following a correct reasoning path towards the solution. We conduct experiments of Step-Back Prompting with PaLM-2 models and observe substantial performance gains on a wide range of challenging reasoning-intensive tasks including STEM, Knowledge QA, and Multi-Hop Reasoning. For instance, Step-Back Prompting improves PaLM-2L performance on MMLU Physics and Chemistry by 7% and 11%, TimeQA by 34%, and MuSiQue by 7%.

We introduce Michelangelo: a minimal, synthetic, and unleaked long-context reasoning evaluation for large language models which is also easy to automatically score. This evaluation is derived via a novel, unifying framework for evaluations over arbitrarily long contexts which measure the model's ability to do more than retrieve a single piece of information from its context. The central idea of the framework () is to construct tasks which require a model to “chisel away” the irrelevant information in the context, revealing a latent structure in the context. To verify a model's understanding of this latent structure, we query the model for details of the structure. Using , we produce three diagnostic long-context evaluations across code and natural-language domains intended to provide a stronger signal of long-context language model capabilities. We perform evaluations on several state-of-the-art models and demonstrate both that a) the proposed evaluations are high-signal and b) that there is significant room for improvement in synthesizing long-context information.

Large language model (LLM) powered chatbots are primarily text-based today, and impose a large interactional cognitive load, especially for exploratory or sensemaking tasks such as planning a trip or learning about a new city. Because the interaction is textual, users have little scaffolding in the way of structure, informational "scent", or ability to specify high-level preferences or goals. We introduce ExploreLLM that allows users to structure thoughts, help explore different options, navigate through the choices and recommendations, and to more easily steer models to generate more personalized responses. We conduct a user study and show that users find it helpful to use ExploreLLM for exploratory or planning tasks, because it provides a useful schema-like structure to the task, and guides users in planning. The study also suggests that users can more easily personalize responses with high-level preferences with ExploreLLM.

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99 as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75 categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

Finetuning language models on a collection of datasets phrased as instructions has been shown to improve model performance and generalization to unseen tasks. In this paper we explore instruction finetuning with a particular focus on (1) scaling the number of tasks, (2) scaling the model size, and (3) finetuning on chain-of-thought data. We find that instruction finetuning with the above aspects dramatically improves performance on a variety of model classes (PaLM, T5, U-PaLM), prompting setups (zero-shot, few-shot, CoT), and evaluation benchmarks (MMLU, BBH, TyDiQA, MGSM, open-ended generation). For instance, Flan-PaLM 540B instruction-finetuned on 1.8K tasks outperforms PALM 540B by a large margin (+9.4% on average). Flan-PaLM 540B achieves state-of-the-art performance on several benchmarks, such as 75.2% on five-shot MMLU. We also publicly release Flan-T5 checkpoints, which achieve strong few-shot performance even compared to much larger models, such as PaLM 62B. Overall, instruction finetuning is a general method for improving the performance and usability of pretrained language models.

Hallucination has been a major problem for large language models and remains a critical challenge when it comes to multimodality in which vision-language models (VLMs) have to deal with not just textual but also visual inputs. Despite rapid progress in VLMs, resources for evaluating and addressing multimodal hallucination are limited and mostly focused on evaluation. This work introduces HaloQuest, a novel visual question answering dataset that captures various aspects of multimodal hallucination such as false premises, insufficient contexts, and visual challenges. A novel idea from HaloQuest is to leverage synthetic images, apart from real ones, to enable dataset creation at scale. With over 7.7K examples spanning across a wide variety of categories, HaloQuest was designed to be both a challenging benchmark for VLMs and a fine-tuning dataset for advancing multimodal reasoning. Our experiments reveal that current models struggle with HaloQuest, with all open-source VLMs achieving below 36% accuracy. On the other hand, fine-tuning on HaloQuest significantly reduces hallucination rates while preserving performance on standard reasoning tasks. Our results discover that benchmarking with generated images is highly correlated (r = 0.97) with real images. Last but not least, we propose a novel Auto-Eval mechanism that is highly correlated with human raters (r = 0.99) for evaluating VLMs. In sum, this work makes concrete strides towards understanding, evaluating, and mitigating hallucination in VLMs, serving as an important step towards more reliable multimodal AI systems in the future.