Jianhao Yan

Jin Xu, Xiaojiang Liu, Jianhao Yan et al. · tencent-ai, tsinghua

While large-scale neural language models, such as GPT2 and BART, have achieved impressive results on various text generation tasks, they tend to get stuck in undesirable sentence-level loops with maximization-based decoding algorithms (\textit{e.g.}, greedy search). This phenomenon is counter-intuitive since there are few consecutive sentence-level repetitions in human corpora (e.g., 0.02\% in Wikitext-103). To investigate the underlying reasons for generating consecutive sentence-level repetitions, we study the relationship between the probabilities of the repetitive tokens and their previous repetitions in the context. Through our quantitative experiments, we find that 1) Language models have a preference to repeat the previous sentence; 2) The sentence-level repetitions have a \textit{self-reinforcement effect}: the more times a sentence is repeated in the context, the higher the probability of continuing to generate that sentence; 3) The sentences with higher initial probabilities usually have a stronger self-reinforcement effect. Motivated by our findings, we propose a simple and effective training method \textbf{DITTO} (Pseu\underline{D}o-Repet\underline{IT}ion Penaliza\underline{T}i\underline{O}n), where the model learns to penalize probabilities of sentence-level repetitions from pseudo repetitive data. Although our method is motivated by mitigating repetitions, experiments show that DITTO not only mitigates the repetition issue without sacrificing perplexity, but also achieves better generation quality. Extensive experiments on open-ended text generation (Wikitext-103) and text summarization (CNN/DailyMail) demonstrate the generality and effectiveness of our method.

Yulong Chen, Huajian Zhang, Yijie Zhou et al. · cambridge, tencent-ai

Most existing cross-lingual summarization (CLS) work constructs CLS corpora by simply and directly translating pre-annotated summaries from one language to another, which can contain errors from both summarization and translation processes. To address this issue, we propose ConvSumX, a cross-lingual conversation summarization benchmark, through a new annotation schema that explicitly considers source input context. ConvSumX consists of 2 sub-tasks under different real-world scenarios, with each covering 3 language directions. We conduct thorough analysis on ConvSumX and 3 widely-used manually annotated CLS corpora and empirically find that ConvSumX is more faithful towards input text. Additionally, based on the same intuition, we propose a 2-Step method, which takes both conversation and summary as input to simulate human annotation process. Experimental results show that 2-Step method surpasses strong baselines on ConvSumX under both automatic and human evaluation. Analysis shows that both source input text and summary are crucial for modeling cross-lingual summaries.

Jianhao Yan, Jin Xu, Fandong Meng et al. · oxford, tencent-ai

Minimum Bayesian Risk Decoding (MBR) emerges as a promising decoding algorithm in Neural Machine Translation. However, MBR performs poorly with label smoothing, which is surprising as label smoothing provides decent improvement with beam search and improves generality in various tasks. In this work, we show that the issue arises from the un-consistency of label smoothing on the token-level and sequence-level distributions. We demonstrate that even though label smoothing only causes a slight change in the token-level, the sequence-level distribution is highly skewed. We coin the issue \emph{autoregressive over-smoothness}. To address this issue, we propose a simple and effective method, Distributional Cooling MBR (DC-MBR), which manipulates the entropy of output distributions by tuning down the Softmax temperature. We theoretically prove the equivalence between pre-tuning label smoothing factor and distributional cooling. Extensive experiments on NMT benchmarks validate that distributional cooling improves MBR in various settings.

Jianhao Yan, Fandong Meng, Jie Zhou · tencent-ai, tsinghua

Hallucination, one kind of pathological translations that bothers Neural Machine Translation, has recently drawn much attention. In simple terms, hallucinated translations are fluent sentences but barely related to source inputs. Arguably, it remains an open problem how hallucination occurs. In this paper, we propose to use probing methods to investigate the causes of hallucinations from the perspective of model architecture, aiming to avoid such problems in future architecture designs. By conducting experiments over various NMT datasets, we find that hallucination is often accompanied by the deficient encoder, especially embeddings, and vulnerable cross-attentions, while, interestingly, cross-attention mitigates some errors caused by the encoder.

Yulong Chen, Yang Liu, Jianhao Yan et al. · cambridge, tencent-ai

The impressive performance of Large Language Models (LLMs) has consistently surpassed numerous human-designed benchmarks, presenting new challenges in assessing the shortcomings of LLMs. Designing tasks and finding LLMs' limitations are becoming increasingly important. In this paper, we investigate the question of whether an LLM can discover its own limitations from the errors it makes. To this end, we propose a Self-Challenge evaluation framework with human-in-the-loop. Starting from seed instances that GPT-4 fails to answer, we prompt GPT-4 to summarize error patterns that can be used to generate new instances and incorporate human feedback on them to refine these patterns for generating more challenging data, iteratively. We end up with 8 diverse patterns, such as text manipulation and questions with assumptions. We then build a benchmark, SC-G4, consisting of 1,835 instances generated by GPT-4 using these patterns, with human-annotated gold responses. The SC-G4 serves as a challenging benchmark that allows for a detailed assessment of LLMs' abilities. Our results show that only 44.96\% of instances in SC-G4 can be answered correctly by GPT-4. Interestingly, our pilot study indicates that these error patterns also challenge other LLMs, such as Claude-3 and Llama-3, and cannot be fully resolved through fine-tuning. Our work takes the first step to demonstrate that LLMs can autonomously identify their inherent flaws and provide insights for future dynamic and automatic evaluation.

Yafu Li, Leyang Cui, Jianhao Yan et al. · tencent-ai, tsinghua

Most existing text generation models follow the sequence-to-sequence paradigm. Generative Grammar suggests that humans generate natural language texts by learning language grammar. We propose a syntax-guided generation schema, which generates the sequence guided by a constituency parse tree in a top-down direction. The decoding process can be decomposed into two parts: (1) predicting the infilling texts for each constituent in the lexicalized syntax context given the source sentence; (2) mapping and expanding each constituent to construct the next-level syntax context. Accordingly, we propose a structural beam search method to find possible syntax structures hierarchically. Experiments on paraphrase generation and machine translation show that the proposed method outperforms autoregressive baselines, while also demonstrating effectiveness in terms of interpretability, controllability, and diversity.

Jianhao Yan, Jin Xu, Chiyu Song et al. · tencent-ai, tsinghua

This paper explores the elusive mechanism underpinning in-context learning in Large Language Models (LLMs). Our work provides a novel perspective by examining in-context learning via the lens of surface repetitions. We quantitatively investigate the role of surface features in text generation, and empirically establish the existence of \emph{token co-occurrence reinforcement}, a principle that strengthens the relationship between two tokens based on their contextual co-occurrences. By investigating the dual impacts of these features, our research illuminates the internal workings of in-context learning and expounds on the reasons for its failures. This paper provides an essential contribution to the understanding of in-context learning and its potential limitations, providing a fresh perspective on this exciting capability.

Jianhao Yan, Pingchuan Yan, Yulong Chen et al. · cambridge, tencent-ai

This study comprehensively evaluates the translation quality of Large Language Models (LLMs), specifically GPT-4, against human translators of varying expertise levels across multiple language pairs and domains. Through carefully designed annotation rounds, we find that GPT-4 performs comparably to junior translators in terms of total errors made but lags behind medium and senior translators. We also observe the imbalanced performance across different languages and domains, with GPT-4's translation capability gradually weakening from resource-rich to resource-poor directions. In addition, we qualitatively study the translation given by GPT-4 and human translators, and find that GPT-4 translator suffers from literal translations, but human translators sometimes overthink the background information. To our knowledge, this study is the first to evaluate LLMs against human translators and analyze the systematic differences between their outputs, providing valuable insights into the current state of LLM-based translation and its potential limitations.

Chiyu Song, Zhanchao Zhou, Jianhao Yan et al.

Instruction tuning is a burgeoning method to elicit the general intelligence of Large Language Models (LLMs). While numerous studies have examined the impact of factors such as data volume and model size on English models, the scaling properties of instruction tuning in other languages remain largely unexplored. In this work, we systematically investigate the effects of data quantity, model size, and data construction methods on instruction tuning for Chinese LLMs. We utilize a newly curated dataset, DoIT, which includes over 40,000 high-quality instruction instances covering ten underlying abilities, such as creative writing, code generation, and logical reasoning. Our experiments, conducted on models ranging from 7b to 33b parameters, yield three key findings: (i) While these factors directly affect overall model performance, some abilities are more responsive to scaling, whereas others demonstrate significant resistance. (ii) The scaling sensitivity of different abilities to these factors can be explained by two features: Complexity and Transference. (iii) By tailoring training strategies to their varying sensitivities, specific abilities can be efficiently learned, enhancing performance on two public benchmarks.

Futing Wang, Jianhao Yan, Yue Zhang et al. · tencent-ai, tsinghua

Enhancing the adaptive capabilities of large language models is a critical pursuit in both research and application. Traditional fine-tuning methods require substantial data and computational resources, especially for enhancing specific capabilities, while in-context learning is limited by the need for appropriate demonstrations and efficient token usage. Inspired by the expression of in-context learned capabilities through task vectors and the concept of modularization, we propose \alg, a framework consisting of two modules designed to effectively store and reuse task vectors to elicit the diverse capabilities of models without additional training or inference tokens. Our comprehensive experiments and analysis demonstrate that our pipeline is highly transferable across different input formats, tasks, and model architectures. ELICIT serves as a plug-and-play performance booster to enable adaptive elicitation of model capabilities. By externally storing and reusing vectors that represent in-context learned capabilities, \alg not only demonstrates the potential to operate modular capabilities but also significantly enhances the performance, versatility, adaptability, and scalability of large language models. Our code will be publicly available at https://github.com/LINs-lab/ELICIT.

Jianhao Yan, Yun Luo, Yue Zhang · tencent-ai, tsinghua

In the multi-turn interaction schema, large language models (LLMs) can leverage user feedback to enhance the quality and relevance of their responses. However, evaluating an LLM's ability to incorporate user refutation feedback is crucial yet challenging. In this study, we introduce RefuteBench 2.0, which significantly extends the original RefuteBench by incorporating LLM agents as refuters and evaluators, which allows for flexible and comprehensive assessment. We design both transient and persistent refutation instructions with different validity periods. Meta-evaluation shows that the LLM-based refuter could generate more human-like refutations and the evaluators could assign scores with high correlation with humans. Experimental results of various LLMs show that current models could effectively satisfy the refutation but fail to memorize the refutation information. Interestingly, we also observe that the performance of the initial task decreases as the refutations increase. Analysis of the attention scores further shows a potential weakness of current LLMs: they struggle to retain and correctly use previous information during long context dialogues. https://github.com/ElliottYan/RefuteBench-2.0

Zican Hu, Shilin Zhang, Yafu Li et al.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a crucial paradigm for incentivizing reasoning capabilities in Large Language Models (LLMs). Due to vast state-action spaces and reward sparsity in reasoning tasks, existing methods often struggle with deficient exploration and poor sample efficiency. In the paper, we propose \textbf{DIVER} (\textbf{D}iversity-\textbf{I}ncentivized Exploration for \textbf{V}ersatil\textbf{E} \textbf{R}easoning), an innovative framework that highlights the pivotal role of global sequence-level diversity to incentivize deep exploration for versatile reasoning. We first conduct a primary empirical study to reveal a strong positive correlation between global diversity and reasoning capacity. Building on this insight, we introduce global diversity incentives as an intrinsic reward to promote deep exploration in a semantically structured space. Incorporating the intrinsic reward, we develop a potential-based reward shaping mechanism to preserve optimal policy invariance and design simple heuristics to mitigate possible reward hacking. Experimental results show that DIVER outperforms competitive RLVR baselines with various exploration strategies on both in-domain and out-of-domain tasks, excelling in both Pass@1 and Pass@k evaluations. Our code is available at https://github.com/NJU-RL/DIVER.

Jianhao Yan, Futing Wang, Yun Luo et al. · tencent-ai, tsinghua

Large language models (LLMs) struggle with maintaining accurate knowledge due to conflicting/outdated parametric memories. While locate-and-edit methods address this, their reliance on models' internal representations leads to robustness failures in long-context reasoning and paraphrased queries. We identify a fundamental limitation of locate-and-edit methods: existing semantic keys (for memory localization) cannot simultaneously satisfy robustness (context-invariant activation) and specificity (precise knowledge discrimination). Through theoretical error-bound analysis, we establish formal criteria for effective editing. Our solution introduces \textit{Robust Edit Pathway (REP)}, a plug-and-play module that: (1) disentangles editing keys from native model representations; (2) dynamically adjusts keys via contrastive learning to achieve robustness-specificity balance. Extensive experiments across various editing methods (ROME/MEMIT/R-ROME/EMMET), existing LLMs (LLaMA2, QWen, Mistral), and datasets (CounterFact, ZsRE) show that REP improves success rate over robustness tests by up-to 66.4\% while maintaining the success rate unaffected. Our code can be found at https://github.com/ElliottYan/RobustKeyEdit .

Xiaoye Qu, Yafu Li, Zhaochen Su et al. · tencent-ai, tsinghua

Recent Large Reasoning Models (LRMs), such as DeepSeek-R1 and OpenAI o1, have demonstrated strong performance gains by scaling up the length of Chain-of-Thought (CoT) reasoning during inference. However, a growing concern lies in their tendency to produce excessively long reasoning traces, which are often filled with redundant content (e.g., repeated definitions), over-analysis of simple problems, and superficial exploration of multiple reasoning paths for harder tasks. This inefficiency introduces significant challenges for training, inference, and real-world deployment (e.g., in agent-based systems), where token economy is critical. In this survey, we provide a comprehensive overview of recent efforts aimed at improving reasoning efficiency in LRMs, with a particular focus on the unique challenges that arise in this new paradigm. We identify common patterns of inefficiency, examine methods proposed across the LRM lifecycle, i.e., from pretraining to inference, and discuss promising future directions for research. To support ongoing development, we also maintain a real-time GitHub repository tracking recent progress in the field. We hope this survey serves as a foundation for further exploration and inspires innovation in this rapidly evolving area.

Jianhao Yan, Yafu Li, Zican Hu et al. · tencent-ai, tsinghua

Recent advances in large reasoning models (LRMs) demonstrate that sophisticated behaviors such as multi-step reasoning and self-reflection can emerge via reinforcement learning with verifiable rewards~(\textit{RLVR}). However, existing \textit{RLVR} approaches are inherently ``on-policy'', limiting learning to a model's own outputs and failing to acquire reasoning abilities beyond its initial capabilities. To address this issue, we introduce \textbf{LUFFY} (\textbf{L}earning to reason \textbf{U}nder o\textbf{FF}-polic\textbf{Y} guidance), a framework that augments \textit{RLVR} with off-policy reasoning traces. LUFFY dynamically balances imitation and exploration by combining off-policy demonstrations with on-policy rollouts during training. Specifically, LUFFY combines the Mixed-Policy GRPO framework, which has a theoretically guaranteed convergence rate, alongside policy shaping via regularized importance sampling to avoid superficial and rigid imitation during mixed-policy training. Compared with previous RLVR methods, LUFFY achieves an over \textbf{+6.4} average gain across six math benchmarks and an advantage of over \textbf{+6.2} points in out-of-distribution tasks. Most significantly, we show that LUFFY successfully trains weak models in scenarios where on-policy RLVR completely fails. These results provide compelling evidence that LUFFY transcends the fundamental limitations of on-policy RLVR and demonstrates the great potential of utilizing off-policy guidance in RLVR.

Jianhao Yan, Futing Wang, Yafu Li et al. · tencent-ai, tsinghua

Large language models (LLMs) trained on vast corpora suffer from inevitable stereotype biases. Mitigating these biases with fine-tuning could be both costly and data-hungry. Model editing methods, which focus on modifying LLMs in a post-hoc manner, are of great potential to address debiasing. However, it lacks a comprehensive study that facilitates both internal and external model editing methods, supports various bias types, as well as understands the pros and cons of applying editing methods to stereotypical debiasing. To mitigate this gap, we carefully formulate social debiasing into an editing problem and benchmark seven existing model editing algorithms on stereotypical debiasing, i.e., debias editing. Our findings in three scenarios reveal both the potential and challenges of debias editing: (1) Existing model editing methods can effectively preserve knowledge and mitigate biases, while the generalization of debias effect from edited sentences to semantically equivalent sentences is limited.(2) Sequential editing highlights the robustness of SERAC (Mitchell et al. 2022b), while internal editing methods degenerate with the number of edits. (3) Model editing algorithms achieve generalization towards unseen biases both within the same type and from different types. In light of these findings, we further propose two simple but effective methods to improve debias editing, and experimentally show the effectiveness of the proposed methods.

Jianhao Yan, Yun Luo, Yue Zhang · tencent-ai, tsinghua

The application scope of large language models (LLMs) is increasingly expanding. In practical use, users might provide feedback based on the model's output, hoping for a responsive model that can complete responses according to their feedback. Whether the model can appropriately respond to users' refuting feedback and consistently follow through with execution has not been thoroughly analyzed. In light of this, this paper proposes a comprehensive benchmark, RefuteBench, covering tasks such as question answering, machine translation, and email writing. The evaluation aims to assess whether models can positively accept feedback in form of refuting instructions and whether they can consistently adhere to user demands throughout the conversation. We conduct evaluations on numerous LLMs and find that LLMs are stubborn, i.e. exhibit inclination to their internal knowledge, often failing to comply with user feedback. Additionally, as the length of the conversation increases, models gradually forget the user's stated feedback and roll back to their own responses. We further propose a recall-and-repeat prompts as a simple and effective way to enhance the model's responsiveness to feedback.

Jianhao Yan, Pingchuan Yan, Yulong Chen et al. · cambridge, tencent-ai

This study presents a comprehensive evaluation of GPT-4's translation capabilities compared to human translators of varying expertise levels. Through systematic human evaluation using the MQM schema, we assess translations across three language pairs (Chinese$\longleftrightarrow$English, Russian$\longleftrightarrow$English, and Chinese$\longleftrightarrow$Hindi) and three domains (News, Technology, and Biomedical). Our findings reveal that GPT-4 achieves performance comparable to junior-level translators in terms of total errors, while still lagging behind senior translators. Unlike traditional Neural Machine Translation systems, which show significant performance degradation in resource-poor language directions, GPT-4 maintains consistent translation quality across all evaluated language pairs. Through qualitative analysis, we identify distinctive patterns in translation approaches: GPT-4 tends toward overly literal translations and exhibits lexical inconsistency, while human translators sometimes over-interpret context and introduce hallucinations. This study represents the first systematic comparison between LLM and human translators across different proficiency levels, providing valuable insights into the current capabilities and limitations of LLM-based translation systems.

Yafu Li, Huajian Zhang, Jianhao Yan et al. · tencent-ai, tsinghua

Recent advances have made non-autoregressive (NAT) translation comparable to autoregressive methods (AT). However, their evaluation using BLEU has been shown to weakly correlate with human annotations. Limited research compares non-autoregressive translation and autoregressive translation comprehensively, leaving uncertainty about the true proximity of NAT to AT. To address this gap, we systematically evaluate four representative NAT methods across various dimensions, including human evaluation. Our empirical results demonstrate that despite narrowing the performance gap, state-of-the-art NAT still underperforms AT under more reliable evaluation metrics. Furthermore, we discover that explicitly modeling dependencies is crucial for generating natural language and generalizing to out-of-distribution sequences.

Jianhao Yan, Chenming Wu, Fandong Meng et al.

Solid evaluation of neural machine translation (NMT) is key to its understanding and improvement. Current evaluation of an NMT system is usually built upon a heuristic decoding algorithm (e.g., beam search) and an evaluation metric assessing similarity between the translation and golden reference. However, this system-level evaluation framework is limited by evaluating only one best hypothesis and search errors brought by heuristic decoding algorithms. To better understand NMT models, we propose a novel evaluation protocol, which defines model errors with model's ranking capability over hypothesis space. To tackle the problem of exponentially large space, we propose two approximation methods, top region evaluation along with an exact top-$k$ decoding algorithm, which finds top-ranked hypotheses in the whole hypothesis space, and Monte Carlo sampling evaluation, which simulates hypothesis space from a broader perspective. To quantify errors, we define our NMT model errors by measuring distance between the hypothesis array ranked by the model and the ideally ranked hypothesis array. After confirming the strong correlation with human judgment, we apply our evaluation to various NMT benchmarks and model architectures. We show that the state-of-the-art Transformer models face serious ranking issues and only perform at the random chance level in the top region. We further analyze model errors on architectures with different depths and widths, as well as different data-augmentation techniques, showing how these factors affect model errors. Finally, we connect model errors with the search algorithms and provide interesting findings of beam search inductive bias and correlation with Minimum Bayes Risk (MBR) decoding.