Multi-agent systems (MAS) have emerged as a powerful paradigm for orchestrating large language models (LLMs) and specialized tools to collaboratively address complex tasks. However, existing MAS frameworks often require manual workflow configuration and lack native support for dynamic evolution and performance optimization. In addition, many MAS optimization algorithms are not integrated into a unified framework. In this paper, we present EvoAgentX, an open-source platform that automates the generation, execution, and evolutionary optimization of multi-agent workflows. EvoAgentX employs a modular architecture consisting of five core layers: the basic components, agent, workflow, evolving, and evaluation layers. Specifically, within the evolving layer, EvoAgentX integrates three MAS optimization algorithms, TextGrad, AFlow, and MIPRO, to iteratively refine agent prompts, tool configurations, and workflow topologies. We evaluate EvoAgentX on HotPotQA, MBPP, and MATH for multi-hop reasoning, code generation, and mathematical problem solving, respectively, and further assess it on real-world tasks using GAIA. Experimental results show that EvoAgentX consistently achieves significant performance improvements, including a 7.44% increase in HotPotQA F1, a 10.00% improvement in MBPP pass@1, a 10.00% gain in MATH solve accuracy, and an overall accuracy improvement of up to 20.00% on GAIA. The source code is available at: https://github.com/EvoAgentX/EvoAgentX
We present any4, a learned 4-bit weight quantization solution for large language models (LLMs) providing arbitrary numeric representations without requiring pre-processing of weights or activations. any4 yields higher accuracy compared to other related 4-bit numeric representation types: int4, fp4 and nf4, as evaluated on a range of model sizes, generations and families (Llama 2, Llama 3, Mistral and Mixtral). While any4 does not require preprocessing of weights or activations, it is also competitive with orthogonal techniques that require such preprocessing (e.g., AWQ and GPTQ). We also experiment with any3 and any2 and show competitiveness at lower bits. Additionally, we show that we can calibrate using a single curated diverse sample rather than hundreds of samples from a dataset as done in most quantization approaches. We also open source tinygemm, a latency optimized GPU matrix multiplication library for LLMs, that implements any4 using a GPU-efficient lookup table strategy along with other common quantization methods. We open source our code at https://github.com/facebookresearch/any4 .
Masked diffusion language models (MDLM) have shown strong promise for non-autoregressive text generation, yet existing samplers act as implicit planners, selecting tokens to unmask via denoiser confidence or entropy scores. Such heuristics falter under parallel unmasking - they ignore pairwise interactions between tokens and cannot account for dependencies when unmasking multiple positions at once, limiting their inference time to traditional auto-regressive (AR) models. We introduce the Dilated-scheduled Unmasking Strategy (DUS), an inference-only, planner-model-free method that requires no additional training. DUS leverages a first-order Markov assumption to partition sequence positions into dilation-based groups of non-adjacent tokens, enabling independent, parallel unmasking steps that respect local context that minimizes the joint entropy of each iteration step. Unlike semi-AR block approaches (e.g., LLADA and Dream) that still invoke the denoiser per block, DUS reduces the number of denoiser calls to O(log B) per generation block - yielding substantial speedup over the O(B) run time of state-of-the-art diffusion models, where B is the block size in the semi-AR inference process. In experiments on math (GSM8K) and code completion (Humaneval, MBPP) benchmarks - domains suited to non-ordinal generation - DUS improves scores over parallel confidence-based planner, without modifying the underlying denoiser. DUS offers a lightweight, budget-aware approach to efficient, high-quality text generation, paving the way to unlock the true capabilities of MDLMs.
Retrieval-Augmented Code Generation (RACG) is a critical technique for enhancing code generation by retrieving relevant information. In this work, we conduct an in-depth analysis of code retrieval by systematically masking specific features while preserving code functionality. Our discoveries include: (1) although trained on code, current retrievers heavily rely on surface-level textual features (e.g., docstrings, identifier names), and (2) they exhibit a strong bias towards well-documented code, even if the documentation is irrelevant.Based on our discoveries, we propose SACL, a framework that enriches textual information and reduces bias by augmenting code or structural knowledge with semantic information. Extensive experiments show that SACL substantially improves code retrieval (e.g., by 12.8% / 9.4% / 7.0% Recall@1 on HumanEval / MBPP / SWE-Bench-Lite), which also leads to better code generation performance (e.g., by 4.88% Pass@1 on HumanEval).
Human reasoning is flexible, adaptive, and grounded in prior experience-qualities that large language models (LLMs) still struggle to emulate. Existing methods either explore diverse reasoning paths at inference time or search for optimal workflows through expensive operations, but both fall short in leveraging multiple reusable strategies in a structured, efficient manner. We propose Guideline Forest, a framework that enhances LLMs reasoning by inducing structured reasoning strategies-called guidelines-from verified examples and executing them via step-wise aggregation. Unlike test-time search or single-path distillation, our method draws on verified reasoning experiences by inducing reusable guidelines and expanding each into diverse variants. Much like human reasoning, these variants reflect alternative thought patterns, are executed in parallel, refined via self-correction, and aggregated step by step-enabling the model to adaptively resolve uncertainty and synthesize robust solutions.We evaluate Guideline Forest on four benchmarks-GSM8K, MATH-500, MBPP, and HumanEval-spanning mathematical and programmatic reasoning. Guideline Forest consistently outperforms strong baselines, including CoT, ReAct, ToT, FoT, and AFlow. Ablation studies further highlight the effectiveness of multi-path reasoning and stepwise aggregation, underscoring the Guideline Forest's adaptability and generalization potential.
Small language models (SLMs) offer promising and efficient alternatives to large language models (LLMs). However, SLMs' limited capacity restricts their reasoning capabilities and makes them sensitive to prompt variations. To address these challenges, we propose a novel framework that enhances SLM reasoning capabilities through LLM generated blueprints. The blueprints provide structured, high-level reasoning guides that help SLMs systematically tackle related problems. Furthermore, our framework integrates a prompt template search mechanism to mitigate the SLMs' sensitivity to prompt variations. Our framework demonstrates improved SLM performance across various tasks, including math (GSM8K), coding (MBPP), and logic reasoning (BBH). Our approach improves the reasoning capabilities of SLMs without increasing model size or requiring additional training, offering a lightweight and deployment-friendly solution for on-device or resource-constrained environments.
Automatic code generation has gained significant momentum with the advent of Large Language Models (LLMs) such as GPT-4. Although many studies focus on improving the effectiveness of LLMs for code generation, very limited work tries to understand the generated code's characteristics and leverage that to improve failed cases. In this paper, as the most straightforward characteristic of code, we investigate the relationship between code complexity and the success of LLM generated code. Using a large set of standard complexity metrics, we first conduct an empirical analysis to explore their correlation with LLM's performance on code generation (i.e., Pass@1). Using logistic regression models, we identify which complexity metrics are most predictive of code correctness. Building on these findings, we propose an iterative feedback method, where LLMs are prompted to generate correct code based on complexity metrics from previous failed outputs. We validate our approach across multiple benchmarks (i.e., HumanEval, MBPP, LeetCode, and BigCodeBench) and various LLMs (i.e., GPT-4o, GPT-3.5 Turbo, Llama 3.1, and GPT-o3 mini), comparing the results with two baseline methods: (a) zero-shot generation, and (b) iterative execution-based feedback without our code complexity insights. Experiment results show that our approach makes notable improvements, particularly with a smaller LLM (GPT3.5 Turbo), where, e.g., Pass@1 increased by 35.71% compared to the baseline's improvement of 12.5% on the HumanEval dataset. The study expands experiments to BigCodeBench and integrates the method with the Reflexion code generation agent, leading to Pass@1 improvements of 20% (GPT-4o) and 23.07% (GPT-o3 mini). The results highlight that complexity-aware feedback enhances both direct LLM prompting and agent-based workflows.
Self-correction has demonstrated potential in code generation by allowing language models to revise and improve their outputs through successive refinement. Recent studies have explored prompting-based strategies that incorporate verification or feedback loops using proprietary models, as well as training-based methods that leverage their strong reasoning capabilities. However, whether smaller models possess the capacity to effectively guide their outputs through self-reflection remains unexplored. Our findings reveal that smaller models struggle to exhibit reflective revision behavior across both self-correction paradigms. In response, we introduce CoCoS, an approach designed to enhance the ability of small language models for multi-turn code correction. Specifically, we propose an online reinforcement learning objective that trains the model to confidently maintain correct outputs while progressively correcting incorrect outputs as turns proceed. Our approach features an accumulated reward function that aggregates rewards across the entire trajectory and a fine-grained reward better suited to multi-turn correction scenarios. This facilitates the model in enhancing initial response quality while achieving substantial improvements through self-correction. With 1B-scale models, CoCoS achieves improvements of 35.8% on the MBPP and 27.7% on HumanEval compared to the baselines.
While Masked Diffusion Models (MDMs), such as LLaDA, present a promising paradigm for language modeling, there has been relatively little effort in aligning these models with human preferences via reinforcement learning. The challenge primarily arises from the high variance in Evidence Lower Bound (ELBO)-based likelihood estimates required for preference optimization. To address this issue, we propose Variance-Reduced Preference Optimization (VRPO), a framework that formally analyzes the variance of ELBO estimators and derives bounds on both the bias and variance of preference optimization gradients. Building on this theoretical foundation, we introduce unbiased variance reduction strategies, including optimal Monte Carlo budget allocation and antithetic sampling, that significantly improve the performance of MDM alignment. We demonstrate the effectiveness of VRPO by applying it to LLaDA, and the resulting model, LLaDA 1.5, outperforms its SFT-only predecessor consistently and significantly across mathematical (GSM8K +4.7), code (HumanEval +3.0, MBPP +1.8), and alignment benchmarks (IFEval +4.0, Arena-Hard +4.3). Furthermore, LLaDA 1.5 demonstrates a highly competitive mathematical performance compared to strong language MDMs and ARMs. Project page: https://ml-gsai.github.io/LLaDA-1.5-Demo/.
Logit-based LLM watermarking traces and verifies AI-generated content by maintaining green and red token lists and increasing the likelihood of green tokens during generation. However, it fails in low-entropy scenarios, where predictable outputs make green token selection difficult without disrupting natural text flow. Existing approaches address this by assuming access to the original LLM to calculate entropy and selectively watermark high-entropy tokens. However, these methods face two major challenges: (1) high computational costs and detection delays due to reliance on the original LLM, and (2) potential risks of model leakage. To address these limitations, we propose Invisible Entropy (IE), a watermarking paradigm designed to enhance both safety and efficiency. Instead of relying on the original LLM, IE introduces a lightweight feature extractor and an entropy tagger to predict whether the entropy of the next token is high or low. Furthermore, based on theoretical analysis, we develop a threshold navigator that adaptively sets entropy thresholds. It identifies a threshold where the watermark ratio decreases as the green token count increases, enhancing the naturalness of the watermarked text and improving detection robustness. Experiments on HumanEval and MBPP datasets demonstrate that IE reduces parameter size by 99\% while achieving performance on par with state-of-the-art methods. Our work introduces a safe and efficient paradigm for low-entropy watermarking. https://github.com/Carol-gutianle/IE https://huggingface.co/datasets/Carol0110/IE-Tagger
We introduce the Diffusion Chain of Lateral Thought (DCoLT), a reasoning framework for diffusion language models. DCoLT treats each intermediate step in the reverse diffusion process as a latent "thinking" action and optimizes the entire reasoning trajectory to maximize the reward on the correctness of the final answer with outcome-based Reinforcement Learning (RL). Unlike traditional Chain-of-Thought (CoT) methods that follow a causal, linear thinking process, DCoLT allows bidirectional, non-linear reasoning with no strict rule on grammatical correctness amid its intermediate steps of thought. We implement DCoLT on two representative Diffusion Language Models (DLMs). First, we choose SEDD as a representative continuous-time discrete diffusion model, where its concrete score derives a probabilistic policy to maximize the RL reward over the entire sequence of intermediate diffusion steps. We further consider the discrete-time masked diffusion language model -- LLaDA, and find that the order to predict and unmask tokens plays an essential role to optimize its RL action resulting from the ranking-based Unmasking Policy Module (UPM) defined by the Plackett-Luce model. Experiments on both math and code generation tasks show that using only public data and 16 H800 GPUs, DCoLT-reinforced DLMs outperform other DLMs trained by SFT or RL or even both. Notably, DCoLT-reinforced LLaDA boosts its reasoning accuracy by +9.8%, +5.7%, +11.4%, +19.5% on GSM8K, MATH, MBPP, and HumanEval.
With the advent of neural language models, the performance of code generation has been significantly boosted. However, the problem of repetitions during the generation process continues to linger. Previous work has primarily focused on content repetition, which is merely a fraction of the broader repetition problem in code generation. A more prevalent and challenging problem is structural repetition. In structural repetition, the repeated code appears in various patterns but possesses a fixed structure, which can be inherently reflected in grammar. In this paper, we formally define structural repetition and propose an efficient decoding approach called RPG, which stands for Repetition Penalization based on Grammar, to alleviate the repetition problems in code generation for LLMs. Specifically, RPG first leverages grammar rules to identify repetition problems during code generation, and then strategically decays the likelihood of critical tokens that contribute to repetitions, thereby mitigating them in code generation. To facilitate this study, we construct a new dataset CodeRepetEval to comprehensively evaluate approaches for mitigating the repetition problems in code generation. Extensive experimental results demonstrate that RPG substantially outperforms the best-performing baselines on CodeRepetEval dataset as well as HumanEval and MBPP benchmarks, effectively reducing repetitions and enhancing the quality of generated code.
The application of large language models (LLMs) in the field of coding is evolving rapidly: from code assistants, to autonomous coding agents, and then to generating complete projects through natural language. Early LLM code benchmarks primarily focused on code generation accuracy, but these benchmarks have gradually become saturated. Benchmark saturation weakens their guiding role for LLMs. For example, HumanEval Pass@1 has reached 99.4% and MBPP 94.2%. Among various attempts to address benchmark saturation, approaches based on software engineering have stood out, but the saturation of existing software engineering benchmarks is rapidly increasing. To address this, we propose a new benchmark, Web-Bench, which contains 50 projects, each consisting of 20 tasks with sequential dependencies. The tasks implement project features in sequence, simulating real-world human development workflows. When designing Web-Bench, we aim to cover the foundational elements of Web development: Web Standards and Web Frameworks. Given the scale and complexity of these projects, which were designed by engineers with 5 to 10 years of experience, each presents a significant challenge. On average, a single project takes 4 to 8 hours for a senior engineer to complete. On our given benchmark agent (Web-Agent), SOTA (Claude 3.7 Sonnet) achieves only 25.1% Pass@1, significantly lower (better) than SWE-Bench's Verified (65.4%) and Full (33.8%) scores. Finally, we discuss that in any development field, Standards and Frameworks represent foundational knowledge and efficiency tools, respectively, and LLMs require optimization tailored to them.
Large Language Models (LLMs) have achieved remarkable success in code generation tasks, powering various applications like code completion, debugging, and programming assistance. However, existing benchmarks such as HumanEval, MBPP, and BigCodeBench primarily evaluate LLMs on English-only prompts, overlooking the real-world scenario where multilingual developers often use code-mixed language while interacting with LLMs. To address this gap, we introduce CodeMixBench, a novel benchmark designed to evaluate the robustness of LLMs on code generation from code-mixed prompts. Built upon BigCodeBench, CodeMixBench introduces controlled code-mixing (CMD) into the natural language parts of prompts across three language pairs: Hinglish (Hindi-English), Spanish-English, and Chinese Pinyin-English. We comprehensively evaluate a diverse set of open-source code generation models ranging from 1.5B to 15B parameters. Our results show that code-mixed prompts consistently degrade Pass@1 performance compared to their English-only counterparts, with performance drops increasing under higher CMD levels for smaller models. CodeMixBench provides a realistic evaluation framework for studying multilingual code generation and highlights new challenges and directions for building robust code generation models that generalize well across diverse linguistic settings.
Multi-agent systems (MAS) built on large language models (LLMs) offer a promising path toward solving complex, real-world tasks that single-agent systems often struggle to manage. While recent advancements in test-time scaling (TTS) have significantly improved single-agent performance on challenging reasoning tasks, how to effectively scale collaboration and reasoning in MAS remains an open question. In this work, we introduce an adaptive multi-agent framework designed to enhance collaborative reasoning through both model-level training and system-level coordination. We construct M500, a high-quality dataset containing 500 multi-agent collaborative reasoning traces, and fine-tune Qwen2.5-32B-Instruct on this dataset to produce M1-32B, a model optimized for multi-agent collaboration. To further enable adaptive reasoning, we propose a novel CEO agent that dynamically manages the discussion process, guiding agent collaboration and adjusting reasoning depth for more effective problem-solving. Evaluated in an open-source MAS across a range of tasks-including general understanding, mathematical reasoning, and coding-our system significantly outperforms strong baselines. For instance, M1-32B achieves 12% improvement on GPQA-Diamond, 41% on AIME2024, and 10% on MBPP-Sanitized, matching the performance of state-of-the-art models like DeepSeek-R1 on some tasks. These results highlight the importance of both learned collaboration and adaptive coordination in scaling multi-agent reasoning. Code is available at https://github.com/jincan333/MAS-TTS
Because large language models are expensive to pretrain on different datasets, using smaller-scale experiments to decide on data is crucial for reducing costs. Which benchmarks and methods of making decisions from observed performance at small scale most accurately predict the datasets that yield the best large models? To empower open exploration of this question, we release models, data, and evaluations in DataDecide -- the most extensive open suite of models over differences in data and scale. We conduct controlled pretraining experiments across 25 corpora with differing sources, deduplication, and filtering up to 100B tokens, model sizes up to 1B parameters, and 3 random seeds. We find that the ranking of models at a single, small size (e.g., 150M parameters) is a strong baseline for predicting best models at our larger target scale (1B) (~80% of com parisons correct). No scaling law methods among 8 baselines exceed the compute-decision frontier of single-scale predictions, but DataDecide can measure improvement in future scaling laws. We also identify that using continuous likelihood metrics as proxies in small experiments makes benchmarks including MMLU, ARC, HellaSwag, MBPP, and HumanEval >80% predictable at the target 1B scale with just 0.01% of the compute.
Large language models (LLMs) have achieved notable success in code generation. However, they still frequently produce uncompilable output because their next-token inference procedure does not model formal aspects of code. Although constrained decoding is a promising approach to alleviate this issue, it has only been applied to handle either domain-specific languages or syntactic features of general-purpose programming languages. However, LLMs frequently generate code with typing errors, which are beyond the domain of syntax and generally hard to adequately constrain. To address this challenge, we introduce a type-constrained decoding approach that leverages type systems to guide code generation. For this purpose, we develop novel prefix automata and a search over inhabitable types, forming a sound approach to enforce well-typedness on LLM-generated code. We formalize our approach on a foundational simply-typed language and extend it to TypeScript to demonstrate practicality. Our evaluation on the HumanEval and MBPP datasets shows that our approach reduces compilation errors by more than half and significantly increases functional correctness in code synthesis, translation, and repair tasks across LLMs of various sizes and model families, including state-of-the-art open-weight models with more than 30B parameters. The results demonstrate the generality and effectiveness of our approach in constraining LLM code generation with formal rules of type systems.
Large Language Models (LLMs) have transformed software development by enabling code generation, automated debugging, and complex reasoning. However, their continued advancement is constrained by the scarcity of high-quality, publicly available supervised fine-tuning (SFT) datasets tailored for coding tasks. To bridge this gap, we introduce OpenCodeInstruct, the largest open-access instruction tuning dataset, comprising 5 million diverse samples. Each sample includes a programming question, solution, test cases, execution feedback, and LLM-generated quality assessments. We fine-tune various base models, including LLaMA and Qwen, across multiple scales (1B+, 3B+, and 7B+) using our dataset. Comprehensive evaluations on popular benchmarks (HumanEval, MBPP, LiveCodeBench, and BigCodeBench) demonstrate substantial performance improvements achieved by SFT with OpenCodeInstruct. We also present a detailed methodology encompassing seed data curation, synthetic instruction and solution generation, and filtering.
The rapid advancement of large language models (LLMs) has significantly improved their performance in code generation tasks. However, existing code benchmarks remain static, consisting of fixed datasets with predefined problems. This makes them vulnerable to memorization during training, where LLMs recall specific test cases instead of generalizing to new problems, leading to data contamination and unreliable evaluation results. To address these issues, we introduce DynaCode, a dynamic, complexity-aware benchmark that overcomes the limitations of static datasets. DynaCode evaluates LLMs systematically using a complexity-aware metric, incorporating both code complexity and call-graph structures. DynaCode achieves large-scale diversity, generating up to 189 million unique nested code problems across four distinct levels of code complexity, referred to as units, and 16 types of call graphs. Results on 12 latest LLMs show an average performance drop of 16.8% to 45.7% compared to MBPP+, a static code generation benchmark, with performance progressively decreasing as complexity increases. This demonstrates DynaCode's ability to effectively differentiate LLMs. Additionally, by leveraging call graphs, we gain insights into LLM behavior, particularly their preference for handling subfunction interactions within nested code.
Code generation and understanding are critical capabilities for large language models (LLMs). Thus, most LLMs are pretrained and fine-tuned on code data. However, these datasets typically treat code as static strings and rarely exploit the dynamic information about their execution. Building upon previous work on trace modeling, we study Execution Tuning (E.T.), a training procedure in which we explicitly model real-world program execution traces without requiring manual test annotations. We train and evaluate models on different execution trace granularities (line and instruction-level) and strategies on the task of output prediction, obtaining around 80% accuracy on CruxEval and MBPP, and showing the advantages of dynamic scratchpads (i.e., self-contained intermediate computations updated by the model rather than accumulated as a history of past computations) on long executions (up to 14k steps). Finally, we discuss E.T.'s practical applications.