Tao Ji

Zhiheng Xi, Dingwen Yang, Jixuan Huang et al.

Training large language models (LLMs) to spend more time thinking and reflection before responding is crucial for effectively solving complex reasoning tasks in fields such as science, coding, and mathematics. However, the effectiveness of mechanisms like self-reflection and self-correction depends on the model's capacity to accurately assess its own performance, which can be limited by factors such as initial accuracy, question difficulty, and the lack of external feedback. In this paper, we delve into a two-player paradigm that separates the roles of reasoning and critique models, where the critique model provides step-level feedback to supervise the reasoning (actor) model during both test-time and train-time. We first propose AutoMathCritique, an automated and scalable framework for collecting critique data, resulting in a dataset of $76,321$ responses paired with step-level feedback. Fine-tuning language models with this dataset enables them to generate natural language feedback for mathematical reasoning. We demonstrate that the critique models consistently improve the actor's performance on difficult queries at test-time, especially when scaling up inference-time computation. Motivated by these findings, we introduce the critique-based supervision to the actor's self-training process, and propose a critique-in-the-loop self-improvement method. Experiments show that the method improves the actor's exploration efficiency and solution diversity, especially on challenging queries, leading to a stronger reasoning model. Lastly, we take the preliminary step to explore training self-talk reasoning models via critique supervision and showcase its potential. Our code and datasets are at \href{https://mathcritique.github.io/}{https://mathcritique.github.io/}.

Xiaoran Fan, Tao Ji, Changhao Jiang et al.

Current large vision-language models (VLMs) often encounter challenges such as insufficient capabilities of a single visual component and excessively long visual tokens. These issues can limit the model's effectiveness in accurately interpreting complex visual information and over-lengthy contextual information. Addressing these challenges is crucial for enhancing the performance and applicability of VLMs. This paper proposes the use of ensemble experts technique to synergizes the capabilities of individual visual encoders, including those skilled in image-text matching, OCR, image segmentation, etc. This technique introduces a fusion network to unify the processing of outputs from different visual experts, while bridging the gap between image encoders and pre-trained LLMs. In addition, we explore different positional encoding schemes to alleviate the waste of positional encoding caused by lengthy image feature sequences, effectively addressing the issue of position overflow and length limitations. For instance, in our implementation, this technique significantly reduces the positional occupancy in models like SAM, from a substantial 4096 to a more efficient and manageable 64 or even down to 1. Experimental results demonstrate that VLMs with multiple experts exhibit consistently superior performance over isolated visual encoders and mark a significant performance boost as more experts are integrated. We have open-sourced the training code used in this report. All of these resources can be found on our project website.

Yi Lu, Xin Zhou, Wei He et al.

Large language models (LLMs) have achieved impressive performance in numerous domains but often struggle to process lengthy inputs effectively and efficiently due to limited length generalization and attention's quadratic computational demands. Many sought to mitigate this by restricting the attention window within the pre-trained length. However, these methods introduce new issues such as ignoring the middle context and requiring additional training. To address these problems, we propose LongHeads, a training-free framework that enhances LLM's long context ability by unlocking multi-head attention's untapped potential. Instead of allowing each head to attend to the full sentence, which struggles with generalizing to longer sequences due to out-of-distribution (OOD) issues, we allow each head to process in-distribution length by selecting and attending to important context chunks. To this end, we propose a chunk selection strategy that relies on the inherent correlation between the query and the key representations, efficiently distributing context chunks to different heads. In this way, each head ensures it can effectively process attended tokens within the trained length, while different heads in different layers can collectively process longer contexts. LongHeads works efficiently in linear time, fits seamlessly with many LLMs that use relative positional encoding. LongHeads achieves 100% accuracy at the 128k length on passkey retrieval task, verifying LongHeads's efficacy in extending the usable context window for existing models. We release our code at https://github.com/LuLuLuyi/LongHeads .

Binghai Wang, Rui Zheng, Lu Chen et al.

Reinforcement Learning from Human Feedback (RLHF) has become a crucial technology for aligning language models with human values and intentions, enabling models to produce more helpful and harmless responses. Reward models are trained as proxies for human preferences to drive reinforcement learning optimization. While reward models are often considered central to achieving high performance, they face the following challenges in practical applications: (1) Incorrect and ambiguous preference pairs in the dataset may hinder the reward model from accurately capturing human intent. (2) Reward models trained on data from a specific distribution often struggle to generalize to examples outside that distribution and are not suitable for iterative RLHF training. In this report, we attempt to address these two issues. (1) From a data perspective, we propose a method to measure the strength of preferences within the data, based on a voting mechanism of multiple reward models. Experimental results confirm that data with varying preference strengths have different impacts on reward model performance. We introduce a series of novel methods to mitigate the influence of incorrect and ambiguous preferences in the dataset and fully leverage high-quality preference data. (2) From an algorithmic standpoint, we introduce contrastive learning to enhance the ability of reward models to distinguish between chosen and rejected responses, thereby improving model generalization. Furthermore, we employ meta-learning to enable the reward model to maintain the ability to differentiate subtle differences in out-of-distribution samples, and this approach can be utilized for iterative RLHF optimization.

Wei Du, Nuowei Liu, Jie Wang et al.

Language models trained with a fixed vocabulary struggle to generalize to novel or out-of-vocabulary words, limiting their flexibility in handling diverse token combinations. Existing dynamic vocabulary approaches attempt to address this limitation but face challenges such as fragmented codebases, lack of support for modern LLMs, and limited inference scalability. To overcome these issues, we introduce DVAGen, a fully open-source, unified framework designed for training, evaluation, and visualization of dynamic vocabulary-augmented language models. Our framework modularizes the pipeline for ease of customization, integrates seamlessly with open-source LLMs, and is the first to provide both CLI and WebUI tools for real-time result inspection. We validate the effectiveness of dynamic vocabulary methods on modern LLMs and demonstrate support for batch inference, significantly improving inference throughput.

Shihan Dou, Yan Liu, Haoxiang Jia et al.

The advancement of large language models (LLMs) has significantly propelled the field of code generation. Previous work integrated reinforcement learning (RL) with compiler feedback for exploring the output space of LLMs to enhance code generation quality. However, the lengthy code generated by LLMs in response to complex human requirements makes RL exploration a challenge. Also, since the unit tests may not cover the complicated code, optimizing LLMs by using these unexecuted code snippets is ineffective. To tackle these challenges, we introduce StepCoder, a novel RL framework for code generation, consisting of two main components: CCCS addresses the exploration challenge by breaking the long sequences code generation task into a Curriculum of Code Completion Subtasks, while FGO only optimizes the model by masking the unexecuted code segments to provide Fine-Grained Optimization. In addition, we furthermore construct the APPS+ dataset for RL training, which is manually verified to ensure the correctness of unit tests. Experimental results show that our method improves the ability to explore the output space and outperforms state-of-the-art approaches in corresponding benchmarks. Our dataset APPS+ and StepCoder are available online.

Jie Wang, Tao Ji, Yuanbin Wu et al.

Generalizing to longer sentences is important for recent Transformer-based language models. Besides algorithms manipulating explicit position features, the success of Transformers without position encodings (NoPE) provides a new way to overcome the challenge. In this paper, we study the length generalization property of NoPE. We find that although NoPE can extend to longer sequences than the commonly used explicit position encodings, it still has a limited context length. We identify a connection between the failure of NoPE's generalization and the distraction of attention distributions. We propose a parameter-efficient tuning for searching attention heads' best temperature hyper-parameters, which substantially expands NoPE's context size. Experiments on long sequence language modeling, the synthetic passkey retrieval task and real-world long context tasks show that NoPE can achieve competitive performances with state-of-the-art length generalization algorithms. The source code is publicly accessible

Tao Ji, Bin Guo, Yuanbin Wu et al.

Multi-head Latent Attention (MLA) is an innovative architecture proposed by DeepSeek, designed to ensure efficient and economical inference by significantly compressing the Key-Value (KV) cache into a latent vector. Compared to MLA, standard LLMs employing Multi-Head Attention (MHA) and its variants such as Grouped-Query Attention (GQA) exhibit significant cost disadvantages. Enabling well-trained LLMs (e.g., Llama) to rapidly adapt to MLA without pre-training from scratch is both meaningful and challenging. This paper proposes the first data-efficient fine-tuning method for transitioning from MHA to MLA (MHA2MLA), which includes two key components: for partial-RoPE, we remove RoPE from dimensions of queries and keys that contribute less to the attention scores, for low-rank approximation, we introduce joint SVD approximations based on the pre-trained parameters of keys and values. These carefully designed strategies enable MHA2MLA to recover performance using only a small fraction (0.3% to 0.6%) of the data, significantly reducing inference costs while seamlessly integrating with compression techniques such as KV cache quantization. For example, the KV cache size of Llama2-7B is reduced by 92.19%, with only a 0.5% drop in LongBench performance.

Shuo Li, Jiajun Sun, Guodong Zheng et al.

Recently, multimodal large language models (MLLMs) have demonstrated remarkable performance in visual-language tasks. However, the authenticity of the responses generated by MLLMs is often compromised by object hallucinations. We identify that a key cause of these hallucinations is the model's over-susceptibility to specific image frequency features in detecting objects. In this paper, we introduce Multi-Frequency Perturbations (MFP), a simple, cost-effective, and pluggable method that leverages both low-frequency and high-frequency features of images to perturb visual feature representations and explicitly suppress redundant frequency-domain features during inference, thereby mitigating hallucinations. Experimental results demonstrate that our method significantly mitigates object hallucinations across various model architectures. Furthermore, as a training-time method, MFP can be combined with inference-time methods to achieve state-of-the-art performance on the CHAIR benchmark.

Nuowei Liu, Jiahao Kuang, Yanting Liu et al.

Protein design is a fundamental challenge in biotechnology, aiming to design novel sequences with specific functions within the vast space of possible proteins. Recent advances in deep generative models have enabled function-based protein design from textual descriptions, yet struggle with structural plausibility. Inspired by classical protein design methods that leverage natural protein structures, we explore whether incorporating fragments from natural proteins can enhance foldability in generative models. Our empirical results show that even random incorporation of fragments improves foldability. Building on this insight, we introduce ProDVa, a novel protein design approach that integrates a text encoder for functional descriptions, a protein language model for designing proteins, and a fragment encoder to dynamically retrieve protein fragments based on textual functional descriptions. Experimental results demonstrate that our approach effectively designs protein sequences that are both functionally aligned and structurally plausible. Compared to state-of-the-art models, ProDVa achieves comparable function alignment using less than 0.04% of the training data, while designing significantly more well-folded proteins, with the proportion of proteins having pLDDT above 70 increasing by 7.38% and those with PAE below 10 increasing by 9.6%.

Xinru Yu, Bin Guo, Shiwei Luo et al.

Causal Language Modeling (CLM) and Masked Language Modeling (MLM) are two mainstream learning paradigms based on Transformer networks, specifically the Decoder-only and Encoder-only architectures. The strengths of each paradigm in downstream tasks have shown a mix of advantages and disadvantages. In the past BabyLM Challenge 2023, although the MLM paradigm achieved the best average performance, the CLM paradigm demonstrated significantly faster convergence rates. For the BabyLM Challenge 2024, we propose a novel language modeling paradigm named $\textbf{AntLM}$, which integrates both CLM and MLM to leverage the advantages of these two classic paradigms. We chose the strict-small track and conducted experiments on two foundation models: BabyLlama, representing CLM, and LTG-BERT, representing MLM. During the training process for specific foundation models, we alternate between applying CLM or MLM training objectives and causal or bidirectional attention masks. Experimental results show that combining the two pretraining objectives leverages their strengths, enhancing overall training performance. Under the same epochs, $AntLM_{BabyLlama}$ improves Macro-average by 1%, and $AntLM_{LTG-BERT}$ achieves a 2.2% increase over the baselines.

Yanting Liu, Tao Ji, Changzhi Sun et al.

We introduce a new dynamic vocabulary for language models. It can involve arbitrary text spans during generation. These text spans act as basic generation bricks, akin to tokens in the traditional static vocabularies. We show that, the ability to generate multi-tokens atomically improve both generation quality and efficiency (compared to the standard language model, the MAUVE metric is increased by 25%, the latency is decreased by 20%). The dynamic vocabulary can be deployed in a plug-and-play way, thus is attractive for various downstream applications. For example, we demonstrate that dynamic vocabulary can be applied to different domains in a training-free manner. It also helps to generate reliable citations in question answering tasks (substantially enhancing citation results without compromising answer accuracy).

Kai Zhong, Luming Sun, Tao Ji et al.

Various works have utilized deep learning to address the query optimization problem in database system. They either learn to construct plans from scratch in a bottom-up manner or steer the plan generation behavior of traditional optimizer using hints. While these methods have achieved some success, they face challenges in either low training efficiency or limited plan search space. To address these challenges, we introduce FOSS, a novel framework for query optimization based on deep reinforcement learning. FOSS initiates optimization from the original plan generated by a traditional optimizer and incrementally refines suboptimal nodes of the plan through a sequence of actions. Additionally, we devise an asymmetric advantage model to evaluate the advantage between two plans. We integrate it with a traditional optimizer to form a simulated environment. Leveraging this simulated environment, FOSS can bootstrap itself to rapidly generate a large amount of high-quality simulated experiences. FOSS then learns from these experiences to improve its optimization capability. We evaluate the performance of FOSS on Join Order Benchmark, TPC-DS, and Stack Overflow. The experimental results demonstrate that FOSS outperforms the state-of-the-art methods in terms of latency performance. Compared to PostgreSQL, FOSS achieves speedup ranging from 1.15x to 8.33x in total latency across different benchmarks.

Liam Davis, Tairan Ji

Modern SMT solvers have revolutionized the approach to constraint satisfaction problems by integrating advanced theory reasoning and encoding techniques. In this work, we evaluate the performance of modern SMT solvers in Z3, CVC5 and DPLL(T) against a standard SAT solver in DPLL. By benchmarking these solvers on novel, diverse 25x25 Sudoku puzzles of various difficulty levels created by our improved Sudoku generator, we examine the impact of advanced theory reasoning and encoding techniques. Our findings demonstrate that modern SMT solvers significantly outperform classical SAT solvers. This work highlights the evolution of logical solvers and exemplifies the utility of SMT solvers in addressing large-scale constraint satisfaction problems.

Chenxiao Liu, Guanzhi Deng, Tao Ji et al.

One challenge for evaluating current sequence- or dialogue-level chatbots, such as Empathetic Open-domain Conversation Models, is to determine whether the chatbot performs in an emotionally consistent way. The most recent work only evaluates on the aspects of context coherence, language fluency, response diversity, or logical self-consistency between dialogues. This work proposes training an evaluator to determine the emotional consistency of chatbots.

Yufang Liu, Tao Ji, Yuanbin Wu et al.

Previous CCG supertaggers usually predict categories using multi-class classification. Despite their simplicity, internal structures of categories are usually ignored. The rich semantics inside these structures may help us to better handle relations among categories and bring more robustness into existing supertaggers. In this work, we propose to generate categories rather than classify them: each category is decomposed into a sequence of smaller atomic tags, and the tagger aims to generate the correct sequence. We show that with this finer view on categories, annotations of different categories could be shared and interactions with sentence contexts could be enhanced. The proposed category generator is able to achieve state-of-the-art tagging (95.5% accuracy) and parsing (89.8% labeled F1) performances on the standard CCGBank. Furthermore, its performances on infrequent (even unseen) categories, out-of-domain texts and low resource language give promising results on introducing generation models to the general CCG analyses.

Tao Ji, Liqian Chen, Xiaoguang Mao et al.

During the software evolution, existing features may be adversely affected by new changes, which is well known as regression errors. Maintaining a high-quality test suite is helpful to prevent regression errors, whereas it heavily depends on developers. Continuously augmenting the existing test suite based on the new changes can assist developers in investigating the impact of these new changes. And by comparing the executions of the generated test case on two versions, existing techniques are able to detect some common errors. However, the requirements and oracles on the new changes with existing program behaviors are missing. In addition, the new changes may introduce new bugs when they are not sufficiently examined with other unchanged code, which finally fails to meet developers' real intentions on changes. In this paper, we propose the notion of evolutionary conflict checking to validate new changes. By extracting developers' intention reflected by new changes and transforming the linear evolutionary process into one three-way merge, we detect conflicts between existing behaviors and new changes. Our experimental results indicate that evolutionary conflict checking is able to be applied for guaranteeing software quality after changes.

Tao Ji, Liqian Chen, Xiaoguang Mao et al.

Merging other branches into the current working branch is common in collaborative software development. However, developers still heavily rely on the textual merge tools to handle the complicated merge tasks. The latent semantic merge conflicts may fail to be detected and degrade the software quality. Regression testing is able to prevent regression faults and has been widely used in real-world software development. However, the merged software may fail to be well examined by rerunning the existing whole test suite. Intuitively, if the test suite fails to cover the changes of different branches at the same time, the merge conflicts would fail to be detected. Recently, it has been proposed to conduct verification on 3-way merges, but this approach does not support even some common cases such as different changes made to different parts of the program. In this paper, we propose an approach of regression unit test generation specifically for checking program merges according to our proposed test oracles. And our general test oracles support us to examine not only 3-way merges, but also 2-way and octopus merges. Considering the conflicts may arise in other locations besides changed methods of the project, we design an algorithm to select UUTs based on the dependency analysis of the whole project. On this basis, we implement a tool called TOM to generate unit tests for Java program merges. We also design the benchmark MCon4J consisting of 389 conflict 3-way merges and 389 conflict octopus merges to facilitate further studies on this topic. The experimental results show that TOM finds 45 conflict 3- way merges and 87 conflicts octopus merges, while the verification based tool fails to work on MCon4J.