Wen-Ding Li

Pengcheng Yin, Wen-Ding Li, Kefan Xiao et al. · stanford

Computational notebooks, such as Jupyter notebooks, are interactive computing environments that are ubiquitous among data scientists to perform data wrangling and analytic tasks. To measure the performance of AI pair programmers that automatically synthesize programs for those tasks given natural language (NL) intents from users, we build ARCADE, a benchmark of 1082 code generation problems using the pandas data analysis framework in data science notebooks. ARCADE features multiple rounds of NL-to-code problems from the same notebook. It requires a model to understand rich multi-modal contexts, such as existing notebook cells and their execution states as well as previous turns of interaction. To establish a strong baseline on this challenging task, we develop PaChiNCo, a 62B code language model (LM) for Python computational notebooks, which significantly outperforms public code LMs. Finally, we explore few-shot prompting strategies to elicit better code with step-by-step decomposition and NL explanation, showing the potential to improve the diversity and explainability of model predictions.

Alexander H. Liu, Alexis Tacnet, Andy Ehrenberg et al. · deepmind, tsinghua

We introduce Voxtral TTS, an expressive multilingual text-to-speech model that generates natural speech from as little as 3 seconds of reference audio. Voxtral TTS adopts a hybrid architecture that combines auto-regressive generation of semantic speech tokens with flow-matching for acoustic tokens. These tokens are encoded and decoded with Voxtral Codec, a speech tokenizer trained from scratch with a hybrid VQ-FSQ quantization scheme. In human evaluations conducted by native speakers, Voxtral TTS is preferred for multilingual voice cloning due to its naturalness and expressivity, achieving a 68.4\% win rate over ElevenLabs Flash v2.5. We release the model weights under a CC BY-NC license.

Alexander H. Liu, Kartik Khandelwal, Sandeep Subramanian et al.

We introduce the Ministral 3 series, a family of parameter-efficient dense language models designed for compute and memory constrained applications, available in three model sizes: 3B, 8B, and 14B parameters. For each model size, we release three variants: a pretrained base model for general-purpose use, an instruction finetuned, and a reasoning model for complex problem-solving. In addition, we present our recipe to derive the Ministral 3 models through Cascade Distillation, an iterative pruning and continued training with distillation technique. Each model comes with image understanding capabilities, all under the Apache 2.0 license.

Kensen Shi, Hanjun Dai, Wen-Ding Li et al.

Search is an important technique in program synthesis that allows for adaptive strategies such as focusing on particular search directions based on execution results. Several prior works have demonstrated that neural models are effective at guiding program synthesis searches. However, a common drawback of those approaches is the inability to handle iterative loops, higher-order functions, or lambda functions, thus limiting prior neural searches from synthesizing longer and more general programs. We address this gap by designing a search algorithm called LambdaBeam that can construct arbitrary lambda functions that compose operations within a given DSL. We create semantic vector representations of the execution behavior of the lambda functions and train a neural policy network to choose which lambdas to construct during search, and pass them as arguments to higher-order functions to perform looping computations. Our experiments show that LambdaBeam outperforms neural, symbolic, and LLM-based techniques in an integer list manipulation domain.

Darren Key, Wen-Ding Li, Kevin Ellis

We develop an approach to estimate the probability that a program sampled from a large language model is correct. Given a natural language description of a programming problem, our method samples both candidate programs as well as candidate predicates specifying how the program should behave. This allows learning a model that forms a well-calibrated probabilistic prediction of program correctness. Our system also infers which predicates are useful to explain the behavior of the generated code, and humans preferred these in a human study over raw language model outputs. Our method is simple, easy to implement, and maintains state of the art generation accuracy results.

Haiming Wang, Mert Unsal, Xiaohan Lin et al. · cambridge

We introduce Kimina-Prover Preview, a large language model that pioneers a novel reasoning-driven exploration paradigm for formal theorem proving, as showcased in this preview release. Trained with a large-scale reinforcement learning pipeline from Qwen2.5-72B, Kimina-Prover demonstrates strong performance in Lean 4 proof generation by employing a structured reasoning pattern we term \textit{formal reasoning pattern}. This approach allows the model to emulate human problem-solving strategies in Lean, iteratively generating and refining proof steps. Kimina-Prover sets a new state-of-the-art on the miniF2F benchmark, reaching 80.7% with pass@8192. Beyond improved benchmark performance, our work yields several key insights: (1) Kimina-Prover exhibits high sample efficiency, delivering strong results even with minimal sampling (pass@1) and scaling effectively with computational budget, stemming from its unique reasoning pattern and RL training; (2) we demonstrate clear performance scaling with model size, a trend previously unobserved for neural theorem provers in formal mathematics; (3) the learned reasoning style, distinct from traditional search algorithms, shows potential to bridge the gap between formal verification and informal mathematical intuition. We open source distilled versions with 1.5B and 7B parameters of Kimina-Prover

Naman Jain, King Han, Alex Gu et al. · berkeley, microsoft-research

Large Language Models (LLMs) applied to code-related applications have emerged as a prominent field, attracting significant interest from both academia and industry. However, as new and improved LLMs are developed, existing evaluation benchmarks (e.g., HumanEval, MBPP) are no longer sufficient for assessing their capabilities. In this work, we propose LiveCodeBench, a comprehensive and contamination-free evaluation of LLMs for code, which continuously collects new problems over time from contests across three competition platforms, namely LeetCode, AtCoder, and CodeForces. Notably, our benchmark also focuses on a broader range of code related capabilities, such as self-repair, code execution, and test output prediction, beyond just code generation. Currently, LiveCodeBench hosts four hundred high-quality coding problems that were published between May 2023 and May 2024. We have evaluated 18 base LLMs and 34 instruction-tuned LLMs on LiveCodeBench. We present empirical findings on contamination, holistic performance comparisons, potential overfitting in existing benchmarks as well as individual model comparisons. We will release all prompts and model completions for further community analysis, along with a general toolkit for adding new scenarios and model

Mistral-AI, Abhinav Rastogi, Albert Q. Jiang et al.

We introduce Magistral, Mistral's first reasoning model and our own scalable reinforcement learning (RL) pipeline. Instead of relying on existing implementations and RL traces distilled from prior models, we follow a ground up approach, relying solely on our own models and infrastructure. Notably, we demonstrate a stack that enabled us to explore the limits of pure RL training of LLMs, present a simple method to force the reasoning language of the model, and show that RL on text data alone maintains most of the initial checkpoint's capabilities. We find that RL on text maintains or improves multimodal understanding, instruction following and function calling. We present Magistral Medium, trained for reasoning on top of Mistral Medium 3 with RL alone, and we open-source Magistral Small (Apache 2.0) which further includes cold-start data from Magistral Medium.

Anton Lozhkov, Raymond Li, Loubna Ben Allal et al. · berkeley, ibm-research

The BigCode project, an open-scientific collaboration focused on the responsible development of Large Language Models for Code (Code LLMs), introduces StarCoder2. In partnership with Software Heritage (SWH), we build The Stack v2 on top of the digital commons of their source code archive. Alongside the SWH repositories spanning 619 programming languages, we carefully select other high-quality data sources, such as GitHub pull requests, Kaggle notebooks, and code documentation. This results in a training set that is 4x larger than the first StarCoder dataset. We train StarCoder2 models with 3B, 7B, and 15B parameters on 3.3 to 4.3 trillion tokens and thoroughly evaluate them on a comprehensive set of Code LLM benchmarks. We find that our small model, StarCoder2-3B, outperforms other Code LLMs of similar size on most benchmarks, and also outperforms StarCoderBase-15B. Our large model, StarCoder2- 15B, significantly outperforms other models of comparable size. In addition, it matches or outperforms CodeLlama-34B, a model more than twice its size. Although DeepSeekCoder- 33B is the best-performing model at code completion for high-resource languages, we find that StarCoder2-15B outperforms it on math and code reasoning benchmarks, as well as several low-resource languages. We make the model weights available under an OpenRAIL license and ensure full transparency regarding the training data by releasing the SoftWare Heritage persistent IDentifiers (SWHIDs) of the source code data.

Long Phan, Alice Gatti, Ziwen Han et al. · amazon-science, apple-ml

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Abhinav Rastogi, Adam Yang, Albert Q. Jiang et al. · deepmind

We introduce Devstral-Small, a lightweight open source model for code agents with the best performance among models below 100B size. In this technical report, we give an overview of how we design and develop a model and craft specializations in agentic software development. The resulting model, Devstral-Small is a small 24B model, fast and easy to serve. Despite its size, Devstral-Small still attains competitive performance compared to models more than an order of magnitude larger.

Kenneth Enevoldsen, Isaac Chung, Imene Kerboua et al. · cambridge, meta-ai

Text embeddings are typically evaluated on a limited set of tasks, which are constrained by language, domain, and task diversity. To address these limitations and provide a more comprehensive evaluation, we introduce the Massive Multilingual Text Embedding Benchmark (MMTEB) - a large-scale, community-driven expansion of MTEB, covering over 500 quality-controlled evaluation tasks across 250+ languages. MMTEB includes a diverse set of challenging, novel tasks such as instruction following, long-document retrieval, and code retrieval, representing the largest multilingual collection of evaluation tasks for embedding models to date. Using this collection, we develop several highly multilingual benchmarks, which we use to evaluate a representative set of models. We find that while large language models (LLMs) with billions of parameters can achieve state-of-the-art performance on certain language subsets and task categories, the best-performing publicly available model is multilingual-e5-large-instruct with only 560 million parameters. To facilitate accessibility and reduce computational cost, we introduce a novel downsampling method based on inter-task correlation, ensuring a diverse selection while preserving relative model rankings. Furthermore, we optimize tasks such as retrieval by sampling hard negatives, creating smaller but effective splits. These optimizations allow us to introduce benchmarks that drastically reduce computational demands. For instance, our newly introduced zero-shot English benchmark maintains a ranking order similar to the full-scale version but at a fraction of the computational cost.

Wen-Ding Li, Keya Hu, Carter Larsen et al.

When learning an input-output mapping from very few examples, is it better to first infer a latent function that explains the examples, or is it better to directly predict new test outputs, e.g. using a neural network? We study this question on ARC by training neural models for induction (inferring latent functions) and transduction (directly predicting the test output for a given test input). We train on synthetically generated variations of Python programs that solve ARC training tasks. We find inductive and transductive models solve different kinds of test problems, despite having the same training problems and sharing the same neural architecture: Inductive program synthesis excels at precise computations, and at composing multiple concepts, while transduction succeeds on fuzzier perceptual concepts. Ensembling them approaches human-level performance on ARC.

Alex Gu, Wen-Ding Li, Naman Jain et al. · berkeley, microsoft-research

While language models are increasingly more proficient at code generation, they still frequently generate incorrect programs. Many of these programs are obviously wrong, but others are more subtle and pass weaker correctness checks such as being able to compile. In this work, we focus on these counterfeit samples: programs sampled from a language model that 1) have a high enough log-probability to be generated at a moderate temperature and 2) pass weak correctness checks. Overall, we discover that most models have a very shallow understanding of counterfeits through three clear failure modes. First, models mistakenly classify them as correct. Second, models are worse at reasoning about the execution behaviour of counterfeits and often predict their execution results as if they were correct. Third, when asking models to fix counterfeits, the likelihood of a model successfully repairing a counterfeit is often even lower than that of sampling a correct program from scratch. Counterfeits also have very unexpected properties: first, counterfeit programs for problems that are easier for a model to solve are not necessarily easier to detect and only slightly easier to execute and repair. Second, counterfeits from a given model are just as confusing to the model itself as they are to other models. Finally, both strong and weak models are able to generate counterfeit samples that equally challenge all models. In light of our findings, we recommend that care and caution be taken when relying on models to understand their own samples, especially when no external feedback is incorporated.

Alex Havrilla, Andrew Dai, Laura O'Mahony et al.

Synthetic data generation with Large Language Models is a promising paradigm for augmenting natural data over a nearly infinite range of tasks. Given this variety, direct comparisons among synthetic data generation algorithms are scarce, making it difficult to understand where improvement comes from and what bottlenecks exist. We propose to evaluate algorithms via the makeup of synthetic data generated by each algorithm in terms of data quality, diversity, and complexity. We choose these three characteristics for their significance in open-ended processes and the impact each has on the capabilities of downstream models. We find quality to be essential for in-distribution model generalization, diversity to be essential for out-of-distribution generalization, and complexity to be beneficial for both. Further, we emphasize the existence of Quality-Diversity trade-offs in training data and the downstream effects on model performance. We then examine the effect of various components in the synthetic data pipeline on each data characteristic. This examination allows us to taxonomize and compare synthetic data generation algorithms through the components they utilize and the resulting effects on data QDC composition. This analysis extends into a discussion on the importance of balancing QDC in synthetic data for efficient reinforcement learning and self-improvement algorithms. Analogous to the QD trade-offs in training data, often there exist trade-offs between model output quality and output diversity which impact the composition of synthetic data. We observe that many models are currently evaluated and optimized only for output quality, thereby limiting output diversity and the potential for self-improvement. We argue that balancing these trade-offs is essential to the development of future self-improvement algorithms and highlight a number of works making progress in this direction.

Junxiong Wang, Wen-Ding Li, Daniele Paliotta et al.

Effective reasoning is crucial to solving complex mathematical problems. Recent large language models (LLMs) have boosted performance by scaling test-time computation through long chain-of-thought reasoning. However, transformer-based models are inherently limited in extending context length due to their quadratic computational complexity and linear memory requirements. In this paper, we introduce a novel hybrid linear RNN reasoning model, M1, built on the Mamba architecture, which allows memory-efficient inference. Our approach leverages a distillation process from existing reasoning models and is further enhanced through RL training. Experimental results on the AIME and MATH benchmarks show that M1 not only outperforms previous linear RNN models but also matches the performance of state-of-the-art Deepseek R1 distilled reasoning models at a similar scale. We also compare our generation speed with a highly performant general purpose inference engine, vLLM, and observe more than a 3x speedup compared to a same size transformer. With throughput speedup, we are able to achieve higher accuracy compared to DeepSeek R1 distilled transformer reasoning models under a fixed generation time budget using self-consistency voting. Overall, we introduce a hybrid Mamba reasoning model and provide a more effective approach to scaling test-time generation using self-consistency or long chain of thought reasoning.

Alex Gu, Naman Jain, Wen-Ding Li et al. · berkeley, microsoft-research

AI for software engineering has made remarkable progress recently, becoming a notable success within generative AI. Despite this, there are still many challenges that need to be addressed before automated software engineering reaches its full potential. It should be possible to reach high levels of automation where humans can focus on the critical decisions of what to build and how to balance difficult tradeoffs while most routine development effort is automated away. Reaching this level of automation will require substantial research and engineering efforts across academia and industry. In this paper, we aim to discuss progress towards this in a threefold manner. First, we provide a structured taxonomy of concrete tasks in AI for software engineering, emphasizing the many other tasks in software engineering beyond code generation and completion. Second, we outline several key bottlenecks that limit current approaches. Finally, we provide an opinionated list of promising research directions toward making progress on these bottlenecks, hoping to inspire future research in this rapidly maturing field.

Alexander H. Liu, Andy Ehrenberg, Andy Lo et al. · deepmind

We present Voxtral Mini and Voxtral Small, two multimodal audio chat models. Voxtral is trained to comprehend both spoken audio and text documents, achieving state-of-the-art performance across a diverse range of audio benchmarks, while preserving strong text capabilities. Voxtral Small outperforms a number of closed-source models, while being small enough to run locally. A 32K context window enables the model to handle audio files up to 40 minutes in duration and long multi-turn conversations. We also contribute three benchmarks for evaluating speech understanding models on knowledge and trivia. Both Voxtral models are released under Apache 2.0 license.

Alexander H. Liu, Andy Ehrenberg, Andy Lo et al.

We introduce Voxtral Realtime, a natively streaming automatic speech recognition model that matches offline transcription quality at sub-second latency. Unlike approaches that adapt offline models through chunking or sliding windows, Voxtral Realtime is trained end-to-end for streaming, with explicit alignment between audio and text streams. Our architecture builds on the Delayed Streams Modeling framework, introducing a new causal audio encoder and Ada RMS-Norm for improved delay conditioning. We scale pretraining to a large-scale dataset spanning 13 languages. At a delay of 480ms, Voxtral Realtime achieves performance on par with Whisper, the most widely deployed offline transcription system. We release the model weights under the Apache 2.0 license.

Terry Yue Zhuo, Minh Chien Vu, Jenny Chim et al.

Task automation has been greatly empowered by the recent advances in Large Language Models (LLMs) via Python code, where the tasks ranging from software engineering development to general-purpose reasoning. While current benchmarks have shown that LLMs can solve tasks using programs like human developers, the majority of their evaluations are limited to short and self-contained algorithmic tasks or standalone function calls. Solving challenging and practical tasks requires the capability of utilizing diverse function calls as tools to efficiently implement functionalities like data analysis and web development. In addition, using multiple tools to solve a task needs compositional reasoning by accurately understanding complex instructions. Fulfilling both of these characteristics can pose a great challenge for LLMs.To assess how well LLMs can solve challenging and practical tasks via programs, we introduce BigCodeBench, a benchmark that challenges LLMs to invoke multiple function calls as tools from 139 libraries and 7 domains for 1,140 fine-grained tasks. To evaluate LLMs rigorously, each task encompasses 5.6 test cases with an average branch coverage of 99%. In addition, we propose a natural-language-oriented variant of BigCodeBench, BigCodeBench-Instruct, that automatically transforms the original docstrings into short instructions only with essential information. Our extensive evaluation of 60 LLMs shows that LLMs are not yet capable of following complex instructions to use function calls precisely, with scores up to 60%, significantly lower than the human performance of 97%. The results underscore the need for further advancements in this area.

Wen-Ding Li, Kevin Ellis

Programming-by-Examples (PBE) aims to generate an algorithm from input-output examples. Such systems are practically and theoretically important: from an end-user perspective, they are deployed to millions of people, and from an AI perspective, PBE corresponds to a very general form of few-shot inductive inference. Given the success of Large Language Models (LLMs) in code-generation tasks, we investigate here the extent to which LLMs can be said to have "solved" PBE. We experiment on classic domains such as lists and strings, and an uncommon graphics programming domain not well represented in typical pretraining data. We find that pretrained models are not effective at PBE, but that they can be fine-tuned for much higher performance, provided the test problems are in-distribution. We analyze empirically what causes these models to succeed and fail, and take steps toward understanding how to achieve better out-of-distribution generalization. Collectively these results suggest that LLMs make strong progress toward solving the typical suite of PBE tasks, potentially increasing the flexibility and applicability of PBE systems, while also identifying ways in which LLMs still fall short.