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 .
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.
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
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.
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
We introduce KodCode, a synthetic dataset that addresses the persistent challenge of acquiring high-quality, verifiable training data across diverse difficulties and domains for training Large Language Models for coding. Existing code-focused resources typically fail to ensure either the breadth of coverage (e.g., spanning simple coding tasks to advanced algorithmic problems) or verifiable correctness (e.g., unit tests). In contrast, KodCode comprises question-solution-test triplets that are systematically validated via a self-verification procedure. Our pipeline begins by synthesizing a broad range of coding questions, then generates solutions and test cases with additional attempts allocated to challenging problems. Finally, post-training data synthesis is done by rewriting questions into diverse formats and generating responses under a test-based reject sampling procedure from a reasoning model (DeepSeek R1). This pipeline yields a large-scale, robust and diverse coding dataset. KodCode is suitable for supervised fine-tuning and the paired unit tests also provide great potential for RL tuning. Fine-tuning experiments on coding benchmarks (HumanEval(+), MBPP(+), BigCodeBench, and LiveCodeBench) demonstrate that KodCode-tuned models achieve state-of-the-art performance, surpassing models like Qwen2.5-Coder-32B-Instruct and DeepSeek-R1-Distill-Llama-70B.
Multi-agent systems (MAS) powered by Large Language Models (LLMs) have been demonstrated to push the boundaries of LLM capabilities, yet they often incur significant costs and face challenges in dynamic LLM selection. Current LLM routing methods effectively reduce overhead in single-agent scenarios by customizing LLM selection for each query, but they overlook the critical decisions regarding collaboration modes and agent roles in MAS. In response to this challenge, we first introduce the problem of Multi-Agent System Routing (MASR), which integrates all components of MAS into a unified routing framework. Toward this goal, we propose MasRouter, the first high-performing, cost-effective, and inductive MASR solution. MasRouter employs collaboration mode determination, role allocation, and LLM routing through a cascaded controller network, progressively constructing a MAS that balances effectiveness and efficiency. Extensive experiments demonstrate that MasRouter is (1) high-performing, achieving a $1.8\%\sim8.2\%$ improvement over the state-of-the-art method on MBPP; (2) economical, reducing overhead by up to $52.07\%$ compared to SOTA methods on HumanEval; and (3) plug-and-play, seamlessly integrating with mainstream MAS frameworks, reducing overhead by $17.21\%\sim28.17\%$ via customized routing. The code is available at https://github.com/yanweiyue/masrouter.
Unit tests (UTs) play an instrumental role in assessing code correctness as well as providing feedback to large language models (LLMs), motivating automated test generation. However, we uncover a trade-off between generating unit test inputs that reveal errors when given a faulty code and correctly predicting the unit test output without access to the gold solution. To address this trade-off, we propose UTGen, which teaches LLMs to generate unit test inputs that reveal errors along with their correct expected outputs based on task descriptions. Since model-generated tests can provide noisy signals (e.g., from incorrectly predicted outputs), we propose UTDebug that (i) scales UTGen via test-time compute to improve UT output prediction, and (ii) validates and backtracks edits based on multiple generated UTs to avoid overfitting, and helps LLMs debug effectively. We show that UTGen outperforms other LLM-based baselines by 7.59% based on a metric measuring the presence of both error-revealing UT inputs and correct UT outputs. When used with UTDebug, we find that feedback from UTGen's unit tests improves pass@1 accuracy of Qwen2.5 32B on HumanEvalFix and our own harder debugging split of MBPP+ by over 3.17% and 12.35% (respectively) over other LLM-based UT generation baselines. Lastly, we demonstrate that UTGen is a better judge for code correctness, outperforming a state-of-the-art trained 8B reward model by 4.43% on HumanEval+ with best-of-10 sampling using Qwen2.5 7B.
Large Language Models (LLMs) demand significant computational resources, making it essential to enhance their capabilities without retraining from scratch. A key challenge in this domain is \textit{catastrophic forgetting} (CF), which hampers performance during Continuous Pre-training (CPT) and Continuous Supervised Fine-Tuning (CSFT). We propose \textbf{Control LLM}, a novel approach that leverages parallel pre-trained and expanded transformer blocks, aligning their hidden-states through interpolation strategies This method effectively preserves performance on existing tasks while seamlessly integrating new knowledge. Extensive experiments demonstrate the effectiveness of Control LLM in both CPT and CSFT. On Llama3.1-8B-Instruct, it achieves significant improvements in mathematical reasoning ($+14.4\%$ on Math-Hard) and coding performance ($+10\%$ on MBPP-PLUS). On Llama3.1-8B, it enhances multilingual capabilities ($+10.6\%$ on C-Eval, $+6.8\%$ on CMMLU, and $+30.2\%$ on CMMLU-0shot-CoT). It surpasses existing methods and achieves SOTA among open-source models tuned from the same base model, using substantially less data and compute. Crucially, these gains are realized while preserving strong original capabilities, with minimal degradation ($<4.3\% \text{on MMLU}$) compared to $>35\%$ in open-source Math and Coding models. This approach has been successfully deployed in LinkedIn's GenAI-powered job seeker and Ads unit products. To support further research, we release the training and evaluation code (https://github.com/linkedin/ControlLLM) along with models trained on public datasets (https://huggingface.co/ControlLLM) to the community.
We introduce self-invoking code generation, a new task designed to evaluate the progressive reasoning and problem-solving capabilities of LLMs. In this task, models are presented with a base problem and a related, more complex problem. They must solve the base problem and then utilize its solution to address the more complex one. This work features three key contributions. First, we propose a general recipe for generating more challenging versions of existing benchmarks, resulting in three new benchmarks: HumanEval Pro, MBPP Pro, and BigCodeBench-Lite Pro, specifically designed to assess LLMs on self-invoking code generation. Second, from the analysis of experimental results over twenty LLMs on our benchmarks, we have two important observations: (i) Most LLMs excel in traditional code generation benchmarks like HumanEval and MBPP, but their performance declines on self-invoking tasks. For example, o1-mini achieves 96.2% pass@1 on HumanEval but only 76.2% on HumanEval Pro. (ii) On self-invoking code generation task, the instruction-tuned models demonstrate only marginal improvements compared to the base models. Third, we disclose the types of failure modes that exist in our evaluation results. All these results underscore the need for further advancements in self-invoking code generation tasks and provide a new direction for future research on enhancing LLMs' code reasoning capabilities.
Large Language Models (LLMs) are widely adopted for assisting in software development tasks, yet their performance evaluations have narrowly focused on the functional correctness of generated code. Human programmers, however, require LLM-generated code to be not only correct but also optimally efficient. We propose PerfCodeGen, a training-free framework that enhances the performance of LLM-generated code by incorporating feedback based on runtime during test case execution into the self-refinement iterations. With PerfCodeGen, we achieve speedups for a significantly higher proportion of problems compared to using the base LLM with sophisticated prompting techniques. Applied to open language models like Phi-3-mini, PerfCodeGen achieves runtime efficiency comparable to prompting powerful closed models like GPT-4. We achieve state-of-the-art runtime efficiency on benchmarks such as HumanEval, MBPP, and APPS, frequently surpassing the ground truth reference solutions with PerfCodeGen using GPT-3.5 and GPT-4. Additionally, we demonstrate the effectiveness of our approach in enhancing code quality across a range of open LLMs of varying sizes including Phi-3-mini, Llama 3 8B, Mixtral 8x7B, Command R, and Llama 3 70B.
Recent research has generated hope that inference scaling could allow weaker language models to match or exceed the accuracy of stronger models, such as by repeatedly sampling solutions to a coding problem until it passes unit tests. The central thesis of this paper is that there is no free lunch for inference scaling: indefinite accuracy improvement through resampling can only be realized if the "verifier" (in this case, a set of unit tests) is perfect. When the verifier is imperfect, as it almost always is in domains such as reasoning or coding (for example, unit tests have imperfect coverage), there is a nonzero probability of false positives: incorrect solutions that pass the verifier. Resampling cannot decrease this probability, so it imposes an upper bound to the accuracy of resampling-based inference scaling even with an infinite compute budget. We find that there is a very strong correlation between the model's single-sample accuracy (i.e. accuracy without unit tests) and its false positive rate on coding benchmarks HumanEval and MBPP, whose unit tests have limited coverage. Therefore, no amount of inference scaling of weaker models can enable them to match the single-sample accuracy of a sufficiently strong model (Fig. 1a). When we consider that false positives have a negative utility compared to abstaining from producing a solution, it bends the inference scaling curve further downward. Empirically, we find that the optimal number of samples can be less than 10 under realistic assumptions (Fig. 1b). Finally, we show that beyond accuracy, false positives may have other undesirable qualities, such as poor adherence to coding style conventions.
The strong performance of large language models (LLMs) raises extensive discussion on their application to code generation. Recent research suggests continuous program refinements through visible tests to improve code generation accuracy in LLMs. However, these methods suffer from LLMs' inefficiency and limited reasoning capacity. In this work, we propose an LLM programming workflow (LPW) designed to improve both initial code generation and subsequent refinements within a structured two-phase workflow. Specifically, the solution generation phase formulates a solution plan, which is then verified through visible tests to specify the intended natural language solution. Subsequently, the code implementation phase drafts an initial code according to the solution plan and its verification. If the generated code fails the visible tests, the plan verification serves as the intended solution to consistently inform the refinement process for correcting bugs. Compared to state-of-the-art methods across various existing LLMs, LPW significantly improves the Pass@1 accuracy by up to 16.4% on well-established text-to-code generation benchmarks. LPW also sets new state-of-the-art Pass@1 accuracy, achieving 98.2% on HumanEval, 84.8% on MBPP, 59.3% on LiveCode, 62.6% on APPS, and 34.7% on CodeContest, using GPT-4o as the backbone. Our code is publicly available at: https://github.com/you68681/lpw
Recently, a number of repository-level code generation benchmarks-such as CoderEval, DevEval, RepoEval, RepoBench, and LongCodeArena-have emerged to evaluate the capabilities of large language models (LLMs) beyond standalone benchmarks like HumanEval and MBPP. Thus, a natural question is, would LLMs have similar performance in real world coding tasks as their performance in these benchmarks? Unfortunately, one cannot answer this question, since these benchmarks consist of short completions, synthetic examples, or focus on limited scale repositories, failing to represent real-world coding tasks. To address these challenges, we create REPOCOD, a Python code-generation benchmark containing complex tasks with realistic dependencies in real-world large projects and appropriate metrics for evaluating source code. It includes 980 whole-function generation tasks from 11 popular projects, 50.8% of which require repository-level context. REPOCOD includes 314 developer-written test cases per instance for better evaluation. We evaluate ten LLMs on REPOCOD and find that none achieves more than 30% pass@1 on REPOCOD, indicating the necessity of building stronger LLMs that can help developers in real-world software development. In addition, we found that retrieval-augmented generation achieves better results than using target function dependencies as context.
Recently, large language models (LLMs) have achieved significant progress in automated code generation. Despite their strong instruction-following capabilities, these models frequently struggled to align with user intent in coding scenarios. In particular, they were hampered by datasets that lacked diversity and failed to address specialized tasks or edge cases. Furthermore, challenges in supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF) led to failures in generating precise, human-intent-aligned code. To tackle these challenges and improve the code generation performance for automated programming systems, we propose Feedback-driven Adaptive Long/short-term memory reinforced Coding Optimization (i.e., FALCON). FALCON is structured into two hierarchical levels. From the global level, long-term memory improves code quality by retaining and applying learned knowledge. At the local level, short-term memory allows for the incorporation of immediate feedback from compilers and AI systems. Additionally, we introduce meta-reinforcement learning with feedback rewards to solve the global-local bi-level optimization problem and enhance the model's adaptability across diverse code generation tasks. Extensive experiments demonstrate that our technique achieves state-of-the-art performance, leading other reinforcement learning methods by more than 4.5 percentage points on the MBPP benchmark and 6.1 percentage points on the Humaneval benchmark. The open-sourced code is publicly available at https://github.com/titurte/FALCON.
Large Language Models (LLMs) have shown incredible potential in code generation tasks, and recent research in prompt engineering have enhanced LLMs' understanding of textual information. However, ensuring the accuracy of generated code often requires extensive testing and validation by programmers. While LLMs can typically generate code based on task descriptions, their accuracy remains limited, especially for complex tasks that require a deeper understanding of both the problem statement and the code generation process. This limitation is primarily due to the LLMs' need to simultaneously comprehend text and generate syntactically and semantically correct code, without having the capability to automatically refine the code. In real-world software development, programmers rarely produce flawless code in a single attempt based on the task description alone, they rely on iterative feedback and debugging to refine their programs. Inspired by this process, we introduce a novel architecture of LLM-based agents for code generation and automatic debugging: Refinement and Guidance Debugging (RGD). The RGD framework is a multi-LLM-based agent debugger that leverages three distinct LLM agents-Guide Agent, Debug Agent, and Feedback Agent. RGD decomposes the code generation task into multiple steps, ensuring a clearer workflow and enabling iterative code refinement based on self-reflection and feedback. Experimental results demonstrate that RGD exhibits remarkable code generation capabilities, achieving state-of-the-art performance with a 9.8% improvement on the HumanEval dataset and a 16.2% improvement on the MBPP dataset compared to the state-of-the-art approaches and traditional direct prompting approaches. We highlight the effectiveness of the RGD framework in enhancing LLMs' ability to generate and refine code autonomously.
The impressive performance of proprietary LLMs like GPT4 in code generation has led to a trend to replicate these capabilities in open-source models through knowledge distillation (e.g. Code Evol-Instruct). However, these efforts often neglect the crucial aspect of response quality, relying heavily on teacher models for direct response distillation. This paradigm, especially for complex instructions, can degrade the quality of synthesized data, compromising the knowledge distillation process. To this end, our study introduces the Adaptive Modular Response Evolution (AMR-Evol) framework, which employs a two-stage process to refine response distillation. The first stage, modular decomposition, breaks down the direct response into more manageable sub-modules. The second stage, adaptive response evolution, automatically evolves the response with the related function modules. Our experiments with three popular code benchmarks (HumanEval, MBPP, and EvalPlus) attest to the superiority of the AMR-Evol framework over baseline response distillation methods. By comparing with the open-source Code LLMs trained on a similar scale of data, we observed performance enhancements: more than +3.0 points on HumanEval-Plus and +1.0 points on MBPP-Plus, which underscores the effectiveness of our framework. Our codes are available at https://github.com/ChiYeungLaw/AMR-Evol.
Reinforcement learning from human feedback (RLHF) is one of the key techniques that helps large language models (LLMs) to follow instructions and provide helpful and harmless responses. While direct policy optimization methods exist, state-of-the-art LLMs adopt RL-based methods (usually PPO) in RLHF to train the policy to generate good responses guided by a reward model learned from preference data. The main challenge of these methods is the inaccuracy of the intermediate reward model, especially in code generation tasks that require long and complex reasoning to score a response. We find that the reliability of the reward model varies across responses assigned with different rewards. This motivates us to filter the samples whose rewards may be unreliable to improve signal-to-noise ratio during policy learning, resulting in Policy Filtration for Proximal Policy Optimization (PF-PPO). To choose a proper policy filtration strategy for a given reward model, the coefficient of determination ($R^2$) between rewards and actual scores on filtered samples serves as a good metrics and helps us find several promising strategies. We provide extensive experiments to validate the effectiveness of PF-PPO in code generation tasks, and find that some variants of PF-PPO are highly effective and achieve new state-of-the-art performance across 7-billion-parameter models on HumanEval, MBPP, and a new and more challenging LeetCode Contest benchmark.