Wei Fu

Zhiyu Mei, Wei Fu, Jiaxuan Gao et al.

The ever-growing complexity of reinforcement learning (RL) tasks demands a distributed system to efficiently generate and process a massive amount of data. However, existing open-source libraries suffer from various limitations, which impede their practical use in challenging scenarios where large-scale training is necessary. In this paper, we present a novel abstraction on the dataflows of RL training, which unifies diverse RL training applications into a general framework. Following this abstraction, we develop a scalable, efficient, and extensible distributed RL system called ReaLlyScalableRL, which allows efficient and massively parallelized training and easy development of customized algorithms. Our evaluation shows that SRL outperforms existing academic libraries, reaching at most 21x higher training throughput in a distributed setting. On learning performance, beyond performing and scaling well on common RL benchmarks with different RL algorithms, SRL can reproduce the same solution in the challenging hide-and-seek environment as reported by OpenAI with up to 5x speedup in wall-clock time. Notably, SRL is the first in the academic community to perform RL experiments at a large scale with over 15k CPU cores. SRL source code is available at: https://github.com/openpsi-project/srl .

Shusheng Xu, Wei Fu, Jiaxuan Gao et al. · tsinghua

Reinforcement Learning from Human Feedback (RLHF) is currently the most widely used method to align large language models (LLMs) with human preferences. Existing RLHF methods can be roughly categorized as either reward-based or reward-free. Novel applications such as ChatGPT and Claude leverage reward-based methods that first learn a reward model and apply actor-critic algorithms, such as Proximal Policy Optimization (PPO). However, in academic benchmarks, state-of-the-art results are often achieved via reward-free methods, such as Direct Preference Optimization (DPO). Is DPO truly superior to PPO? Why does PPO perform poorly on these benchmarks? In this paper, we first conduct both theoretical and empirical studies on the algorithmic properties of DPO and show that DPO may have fundamental limitations. Moreover, we also comprehensively examine PPO and reveal the key factors for the best performances of PPO in fine-tuning LLMs. Finally, we benchmark DPO and PPO across a collection of RLHF testbeds, ranging from dialogue to code generation. Experiment results demonstrate that PPO is able to surpass other alignment methods in all cases and achieve state-of-the-art results in challenging code competitions. Our code is publicly available at https://github.com/openpsi-project/ReaLHF.

Zhiyu Mei, Wei Fu, Kaiwei Li et al.

Reinforcement Learning from Human Feedback (RLHF) is a pivotal technique for empowering large language model (LLM) applications. Compared with the supervised training process of LLMs, the RLHF training process is much more sophisticated, requiring a diverse range of computation workloads with intricate dependencies between multiple LLM instances. Therefore, simply adopting the fixed parallelization strategies from supervised training for LLMs can be insufficient for RLHF and result in low training efficiency. To overcome this limitation, we propose a novel technique named parameter ReaLlocation, which dynamically adapts the parallelization strategies for different workloads during training by redistributing LLM parameters across the training cluster. Building upon this idea, we introduce ReaL, a pioneering system for efficient RLHF training. ReaL introduces the concept of an execution plan, which defines a fine-grained resource allocation and parallelization strategy particularly designed for RLHF training. Based on this concept, ReaL employs a tailored search algorithm with a lightweight run-time estimator to automatically discover an efficient execution plan for an instance of RLHF experiment. Subsequently, the runtime engine deploys the selected plan by effectively parallelizing computations and redistributing parameters. We evaluate ReaL on the LLaMA models with up to 70 billion parameters and 128 GPUs. The experimental results demonstrate that ReaL achieves speedups of up to $3.58\times$ compared to baseline methods. Furthermore, the execution plans generated by ReaL exhibit an average of $81\%$ performance improvement over heuristic approaches based on Megatron-LM in the long-context scenario. The source code of ReaL is publicly available at https://github.com/openpsi-project/ReaLHF .

Wei Fu, Tim Menzies, Xipeng Shen

Context: Data miners have been widely used in software engineering to, say, generate defect predictors from static code measures. Such static code defect predictors perform well compared to manual methods, and they are easy to use and useful to use. But one of the "black art" of data mining is setting the tunings that control the miner. Objective:We seek simple, automatic, and very effective method for finding those tunings. Method: For each experiment with different data sets (from open source JAVA systems), we ran differential evolution as anoptimizer to explore the tuning space (as a first step) then tested the tunings using hold-out data. Results: Contrary to our prior expectations, we found these tunings were remarkably simple: it only required tens, not thousands,of attempts to obtain very good results. For example, when learning software defect predictors, this method can quickly find tuningsthat alter detection precision from 0% to 60%. Conclusion: Since (1) the improvements are so large, and (2) the tuning is so simple, we need to change standard methods insoftware analytics. At least for defect prediction, it is no longer enough to just run a data miner and present the resultwithoutconducting a tuning optimization study. The implication for other kinds of analytics is now an open and pressing issue

Jiaxuan Gao, Shusheng Xu, Wenjie Ye et al. · tsinghua

Reward models have been increasingly critical for improving the reasoning capability of LLMs. Existing research has shown that a well-trained reward model can substantially improve model performances at inference time via search. However, the potential of reward models during RL training time still remains largely under-explored. It is currently unclear whether these reward models can provide additional training signals to enhance the reasoning capabilities of LLMs in RL training that uses sparse success rewards, which verify the correctness of solutions. In this work, we evaluate popular reward models for RL training, including the Outcome-supervised Reward Model (ORM) and the Process-supervised Reward Model (PRM), and train a collection of LLMs for math problems using RL by combining these learned rewards with success rewards. Surprisingly, even though these learned reward models have strong inference-time performances, they may NOT help or even hurt RL training, producing worse performances than LLMs trained with the success reward only. Our analysis reveals that an LLM can receive high rewards from some of these reward models by repeating correct but unnecessary reasoning steps, leading to a severe reward hacking issue. Therefore, we introduce two novel reward refinement techniques, including Clipping and Delta. The key idea is to ensure the accumulative reward of any reasoning trajectory is upper-bounded to keep a learned reward model effective without being exploited. We evaluate our techniques with multiple reward models over a set of 1.5B and 7B LLMs on MATH and GSM8K benchmarks and demonstrate that with a carefully designed reward function, RL training without any additional supervised tuning can improve all the evaluated LLMs, including the state-of-the-art 7B LLM Qwen2.5-Math-7B-Instruct on MATH and GSM8K benchmarks.

Amritanshu Agrawal, Wei Fu, Di Chen et al.

Machine learning techniques applied to software engineering tasks can be improved by hyperparameter optimization, i.e., automatic tools that find good settings for a learner's control parameters. We show that such hyperparameter optimization can be unnecessarily slow, particularly when the optimizers waste time exploring "redundant tunings"', i.e., pairs of tunings which lead to indistinguishable results. By ignoring redundant tunings, DODGE, a tuning tool, runs orders of magnitude faster, while also generating learners with more accurate predictions than seen in prior state-of-the-art approaches.

Di Chen, Wei Fu, Rahul Krishna et al.

Actionable analytics are those that humans can understand, and operationalize. What kind of data mining models generate such actionable analytics? According to psychological scientists, humans understand models that most match their own internal models, which they characterize as lists of "heuristic" (i.e., lists of very succinct rules). One such heuristic rule generator is the Fast-and-Frugal Trees (FFT) preferred by psychological scientists. Despite their successful use in many applied domains, FFTs have not been applied in software analytics. Accordingly, this paper assesses FFTs for software analytics. We find that FFTs are remarkably effective. Their models are very succinct (5 lines or less describing a binary decision tree). These succinct models outperform state-of-the-art defect prediction algorithms defined by Ghortra et al. at ICSE'15. Also, when we restrict training data to operational attributes (i.e., those attributes that are frequently changed by developers), FFTs perform much better than standard learners. Our conclusions are two-fold. Firstly, there is much that software analytics community could learn from psychological science. Secondly, proponents of complex methods should always baseline those methods against simpler alternatives. For example, FFTs could be used as a standard baseline learner against which other software analytics tools are compared.

Wei Fu, Tim Menzies, Di Chen et al.

Despite extensive research, many methods in software quality prediction still exhibit some degree of uncertainty in their results. Rather than treating this as a problem, this paper asks if this uncertainty is a resource that can simplify software quality prediction. For example, Deb's principle of $ε$-dominance states that if there exists some $ε$ value below which it is useless or impossible to distinguish results, then it is superfluous to explore anything less than $ε$. We say that for "large $ε$ problems", the results space of learning effectively contains just a few regions. If many learners are then applied to such large $ε$ problems, they would exhibit a "many roads lead to Rome" property; i.e., many different software quality prediction methods would generate a small set of very similar results. This paper explores DART, an algorithm especially selected to succeed for large $ε$ software quality prediction problems. DART is remarkable simple yet, on experimentation, it dramatically out-performs three sets of state-of-the-art defect prediction methods. The success of DART for defect prediction begs the questions: how many other domains in software quality predictors can also be radically simplified? This will be a fruitful direction for future work.

Suvodeep Majumder, Nikhila Balaji, Katie Brey et al.

Deep learning methods are useful for high-dimensional data and are becoming widely used in many areas of software engineering. Deep learners utilizes extensive computational power and can take a long time to train-- making it difficult to widely validate and repeat and improve their results. Further, they are not the best solution in all domains. For example, recent results show that for finding related Stack Overflow posts, a tuned SVM performs similarly to a deep learner, but is significantly faster to train. This paper extends that recent result by clustering the dataset, then tuning very learners within each cluster. This approach is over 500 times faster than deep learning (and over 900 times faster if we use all the cores on a standard laptop computer). Significantly, this faster approach generates classifiers nearly as good (within 2\% F1 Score) as the much slower deep learning method. Hence we recommend this faster methods since it is much easier to reproduce and utilizes far fewer CPU resources. More generally, we recommend that before researchers release research results, that they compare their supposedly sophisticated methods against simpler alternatives (e.g applying simpler learners to build local models).

Wei Fu, Vivek Nair, Tim Menzies

Context: One of the black arts of data mining is learning the magic parameters which control the learners. In software analytics, at least for defect prediction, several methods, like grid search and differential evolution (DE), have been proposed to learn these parameters, which has been proved to be able to improve the performance scores of learners. Objective: We want to evaluate which method can find better parameters in terms of performance score and runtime cost. Methods: This paper compares grid search to differential evolution, which is an evolutionary algorithm that makes extensive use of stochastic jumps around the search space. Results: We find that the seemingly complete approach of grid search does no better, and sometimes worse, than the stochastic search. When repeated 20 times to check for conclusion validity, DE was over 210 times faster than grid search to tune Random Forests on 17 testing data sets with F-Measure Conclusions: These results are puzzling: why does a quick partial search be just as effective as a much slower, and much more, extensive search? To answer that question, we turned to the theoretical optimization literature. Bergstra and Bengio conjecture that grid search is not more effective than more randomized searchers if the underlying search space is inherently low dimensional. This is significant since recent results show that defect prediction exhibits very low intrinsic dimensionality-- an observation that explains why a fast method like DE may work as well as a seemingly more thorough grid search. This suggests, as a future research direction, that it might be possible to peek at data sets before doing any optimization in order to match the optimization algorithm to the problem at hand.

Amritanshu Agrawal, Wei Fu, Tim Menzies

Context: Topic modeling finds human-readable structures in unstructured textual data. A widely used topic modeler is Latent Dirichlet allocation. When run on different datasets, LDA suffers from "order effects" i.e. different topics are generated if the order of training data is shuffled. Such order effects introduce a systematic error for any study. This error can relate to misleading results;specifically, inaccurate topic descriptions and a reduction in the efficacy of text mining classification results. Objective: To provide a method in which distributions generated by LDA are more stable and can be used for further analysis. Method: We use LDADE, a search-based software engineering tool that tunes LDA's parameters using DE (Differential Evolution). LDADE is evaluated on data from a programmer information exchange site (Stackoverflow), title and abstract text of thousands ofSoftware Engineering (SE) papers, and software defect reports from NASA. Results were collected across different implementations of LDA (Python+Scikit-Learn, Scala+Spark); across different platforms (Linux, Macintosh) and for different kinds of LDAs (VEM,or using Gibbs sampling). Results were scored via topic stability and text mining classification accuracy. Results: In all treatments: (i) standard LDA exhibits very large topic instability; (ii) LDADE's tunings dramatically reduce cluster instability; (iii) LDADE also leads to improved performances for supervised as well as unsupervised learning. Conclusion: Due to topic instability, using standard LDA with its "off-the-shelf" settings should now be depreciated. Also, in future, we should require SE papers that use LDA to test and (if needed) mitigate LDA topic instability. Finally, LDADE is a candidate technology for effectively and efficiently reducing that instability.