In this white paper, we present AlphaEvolve, an evolutionary coding agent that substantially enhances capabilities of state-of-the-art LLMs on highly challenging tasks such as tackling open scientific problems or optimizing critical pieces of computational infrastructure. AlphaEvolve orchestrates an autonomous pipeline of LLMs, whose task is to improve an algorithm by making direct changes to the code. Using an evolutionary approach, continuously receiving feedback from one or more evaluators, AlphaEvolve iteratively improves the algorithm, potentially leading to new scientific and practical discoveries. We demonstrate the broad applicability of this approach by applying it to a number of important computational problems. When applied to optimizing critical components of large-scale computational stacks at Google, AlphaEvolve developed a more efficient scheduling algorithm for data centers, found a functionally equivalent simplification in the circuit design of hardware accelerators, and accelerated the training of the LLM underpinning AlphaEvolve itself. Furthermore, AlphaEvolve discovered novel, provably correct algorithms that surpass state-of-the-art solutions on a spectrum of problems in mathematics and computer science, significantly expanding the scope of prior automated discovery methods (Romera-Paredes et al., 2023). Notably, AlphaEvolve developed a search algorithm that found a procedure to multiply two $4 \times 4$ complex-valued matrices using $48$ scalar multiplications; offering the first improvement, after 56 years, over Strassen's algorithm in this setting. We believe AlphaEvolve and coding agents like it can have a significant impact in improving solutions of problems across many areas of science and computation.

Generating rational and generally accurate responses to tasks, often accompanied by example demonstrations, highlights Large Language Model's (LLM's) remarkable In-Context Learning (ICL) capabilities without requiring updates to the model's parameter space. Despite having an ongoing exploration focused on the inference from a document-level concept, its behavior in learning well-defined functions or relations in context needs a careful investigation. In this article, we present the performance of ICL on partially ordered relation by introducing the notion of inductively increasing complexity in prompts. In most cases, the saturated performance of the chosen metric indicates that while ICL offers some benefits, its effectiveness remains constrained as we increase the complexity in the prompts even in presence of sufficient demonstrative examples. The behavior is evident from our empirical findings and has further been theoretically justified in term of its implicit optimization process. The code is available here.

In this paper, we describe our approach for the SemEval 2025 Task 2 on Entity-Aware Machine Translation (EA-MT). Our system aims to improve the accuracy of translating named entities by combining two key approaches: Retrieval Augmented Generation (RAG) and iterative self-refinement techniques using Large Language Models (LLMs). A distinctive feature of our system is its self-evaluation mechanism, where the LLM assesses its own translations based on two key criteria: the accuracy of entity translations and overall translation quality. We demonstrate how these methods work together and effectively improve entity handling while maintaining high-quality translations.

Training Transformers on algorithmic tasks frequently demonstrates an intriguing abrupt learning phenomenon: an extended performance plateau followed by a sudden, sharp improvement. This work investigates the underlying mechanisms for such dynamics, primarily in shallow Transformers. We reveal that during the plateau, the model often develops an interpretable partial solution while simultaneously exhibiting a strong repetition bias in their outputs. This output degeneracy is accompanied by internal representation collapse, where hidden states across different tokens become nearly parallel. We further identify the slow learning of optimal attention maps as a key bottleneck. Hidden progress in attention configuration during the plateau precedes the eventual rapid convergence, and directly intervening on attention significantly alters plateau duration and the severity of repetition bias and representational collapse. We validate that these identified phenomena-repetition bias and representation collapse-are not artifacts of toy setups but also manifest in the early pre-training stage of large language models like Pythia and OLMo.

Mathematical reasoning presents a significant challenge for Large Language Models (LLMs), often requiring robust multi step logical consistency. While Chain of Thought (CoT) prompting elicits reasoning steps, it doesn't guarantee correctness, and improving reliability via extensive sampling is computationally costly. This paper introduces the Energy Outcome Reward Model (EORM), an effective, lightweight, post hoc verifier. EORM leverages Energy Based Models (EBMs) to simplify the training of reward models by learning to assign a scalar energy score to CoT solutions using only outcome labels, thereby avoiding detailed annotations. It achieves this by interpreting discriminator output logits as negative energies, effectively ranking candidates where lower energy is assigned to solutions leading to correct final outcomes implicitly favoring coherent reasoning. On mathematical benchmarks (GSM8k, MATH), EORM significantly improves final answer accuracy (e.g., with Llama 3 8B, achieving 90.7% on GSM8k and 63.7% on MATH). EORM effectively leverages a given pool of candidate solutions to match or exceed the performance of brute force sampling, thereby enhancing LLM reasoning outcome reliability through its streamlined post hoc verification process.

Inorganic synthesis planning currently relies primarily on heuristic approaches or machine-learning models trained on limited datasets, which constrains its generality. We demonstrate that language models, without task-specific fine-tuning, can recall synthesis conditions. Off-the-shelf models, such as GPT-4.1, Gemini 2.0 Flash and Llama 4 Maverick, achieve a Top-1 precursor-prediction accuracy of up to 53.8 % and a Top-5 performance of 66.1 % on a held-out set of 1,000 reactions. They also predict calcination and sintering temperatures with mean absolute errors below 126 °C, matching specialized regression methods. Ensembling these language models further enhances predictive accuracy and reduces inference cost per prediction by up to 70 %. We subsequently employ language models to generate 28,548 synthetic reaction recipes, which we combine with literature-mined examples to pretrain a transformer-based model, SyntMTE. After fine-tuning on the combined dataset, SyntMTE reduces mean-absolute error in sintering temperature prediction to 73 °C and in calcination temperature to 98 °C. This strategy improves models by up to 8.7 % compared with baselines trained exclusively on experimental data. Finally, in a case study on Li7La3Zr2O12 solid-state electrolytes, we demonstrate that SyntMTE reproduces the experimentally observed dopant-dependent sintering trends. Our hybrid workflow enables scalable, data-efficient inorganic synthesis planning.

Machine unlearning aims to remove specific information, e.g. sensitive or undesirable content, from large language models (LLMs) while preserving overall performance. We propose an inference-time unlearning algorithm that uses contrastive decoding, leveraging two auxiliary smaller models, one trained without the forget set and one trained with it, to guide the outputs of the original model using their difference during inference. Our strategy substantially improves the tradeoff between unlearning effectiveness and model utility. We evaluate our approach on two unlearning benchmarks, TOFU and MUSE. Results show notable gains in both forget quality and retained performance in comparison to prior approaches, suggesting that incorporating contrastive decoding can offer an efficient, practical avenue for unlearning concepts in large-scale models.

The growing number of pretrained models in Machine Learning (ML) presents significant challenges for practitioners. Given a new dataset, they need to determine the most suitable deep learning (DL) pipeline, consisting of the pretrained model and the hyperparameters for finetuning to it. Moreover, as models grow in scale, the increasing reliance on real-world data poses a bottleneck for training and requires leveraging data more effectively. Addressing the first challenge often involves manual model selection and hyperparameter tuning. At the same time, as models grow larger and more and more of the available human-generated data is being used for training, data augmentation and synthetic data become critical elements. Automated machine learning offers a path to address these challenges but is traditionally designed for tabular data and classical ML methods. This dissertation adopts meta-learning to extend automated machine learning to the deep learning domain. We propose empirical approaches to automate DL pipeline selection for Computer Vision tasks using prior task knowledge to learn surrogate models for pipeline ranking. Extending these methods to the language domain, we learn to finetune large language models. As a result, we show that our approach can outperform finetuning foundation models. Additionally, we meta-learn data augmentation and synthetic data to enhance performance in up-stream and down-stream tasks. We empirically show the underestimated importance of data augmentation when using Self-Supervised Learning and meta-learn advanced data augmentation strategies. Leveraging synthetic data, we also propose to meta-learn neural synthetic data generators as proxies for Reinforcement Learning (RL) environments. Additionally, we learn a multiple-environment world model in an in-context learning fashion by purely using synthetic, randomly sampled data.

Preference-based Reinforcement Learning (PbRL) entails a variety of approaches for aligning models with human intent to alleviate the burden of reward engineering. However, most previous PbRL work has not investigated the robustness to labeler errors, inevitable with labelers who are non-experts or operate under time constraints. Additionally, PbRL algorithms often target very specific settings (e.g. pairwise ranked preferences or purely offline learning). We introduce Similarity as Reward Alignment (SARA), a simple contrastive framework that is both resilient to noisy labels and adaptable to diverse feedback formats and training paradigms. SARA learns a latent representation of preferred samples and computes rewards as similarities to the learned latent. We demonstrate strong performance compared to baselines on continuous control offline RL benchmarks. We further demonstrate SARA's versatility in applications such as trajectory filtering for downstream tasks, cross-task preference transfer, and reward shaping in online learning.

Scaling test-time compute has emerged as an effective strategy for improving the performance of large language models. However, existing methods typically allocate compute uniformly across all queries, overlooking variation in query difficulty. To address this inefficiency, we formulate test-time compute allocation as a novel bandit learning problem and propose adaptive algorithms that estimate query difficulty on the fly and allocate compute accordingly. Compared to uniform allocation, our algorithms allocate more compute to challenging queries while maintaining accuracy on easier ones. Among challenging queries, our algorithms further learn to prioritize solvable instances, effectively reducing excessive computing on unsolvable queries. We theoretically prove that our algorithms achieve better compute efficiency than uniform allocation and empirically validate their effectiveness on math and code benchmarks. Specifically, our algorithms achieve up to an 11.10% performance improvement (15.04% relative) on the MATH-500 dataset and up to a 7.41% performance improvement (14.40% relative) on LiveCodeBench.