Runzhe Zhan

Junchao Wu, Shu Yang, Runzhe Zhan et al.

The powerful ability to understand, follow, and generate complex language emerging from large language models (LLMs) makes LLM-generated text flood many areas of our daily lives at an incredible speed and is widely accepted by humans. As LLMs continue to expand, there is an imperative need to develop detectors that can detect LLM-generated text. This is crucial to mitigate potential misuse of LLMs and safeguard realms like artistic expression and social networks from harmful influence of LLM-generated content. The LLM-generated text detection aims to discern if a piece of text was produced by an LLM, which is essentially a binary classification task. The detector techniques have witnessed notable advancements recently, propelled by innovations in watermarking techniques, statistics-based detectors, neural-base detectors, and human-assisted methods. In this survey, we collate recent research breakthroughs in this area and underscore the pressing need to bolster detector research. We also delve into prevalent datasets, elucidating their limitations and developmental requirements. Furthermore, we analyze various LLM-generated text detection paradigms, shedding light on challenges like out-of-distribution problems, potential attacks, real-world data issues and the lack of effective evaluation framework. Conclusively, we highlight interesting directions for future research in LLM-generated text detection to advance the implementation of responsible artificial intelligence (AI). Our aim with this survey is to provide a clear and comprehensive introduction for newcomers while also offering seasoned researchers a valuable update in the field of LLM-generated text detection. The useful resources are publicly available at: https://github.com/NLP2CT/LLM-generated-Text-Detection.

Xinyi Yang, Runzhe Zhan, Derek F. Wong et al.

The large language model (LLM) has garnered significant attention due to its in-context learning mechanisms and emergent capabilities. The research community has conducted several pilot studies to apply LLMs to machine translation tasks and evaluate their performance from diverse perspectives. However, previous research has primarily focused on the LLM itself and has not explored human intervention in the inference process of LLM. The characteristics of LLM, such as in-context learning and prompt engineering, closely mirror human cognitive abilities in language tasks, offering an intuitive solution for human-in-the-loop generation. In this study, we propose a human-in-the-loop pipeline that guides LLMs to produce customized outputs with revision instructions. The pipeline initiates by prompting the LLM to produce a draft translation, followed by the utilization of automatic retrieval or human feedback as supervision signals to enhance the LLM's translation through in-context learning. The human-machine interactions generated in this pipeline are also stored in an external database to expand the in-context retrieval database, enabling us to leverage human supervision in an offline setting. We evaluate the proposed pipeline using GPT-3.5-turbo API on five domain-specific benchmarks for German-English translation. The results demonstrate the effectiveness of the pipeline in tailoring in-domain translations and improving translation performance compared to direct translation. Additionally, we discuss the results from the following perspectives: 1) the effectiveness of different in-context retrieval methods; 2) the construction of a retrieval database under low-resource scenarios; 3) the observed domains differences; 4) the quantitative analysis of linguistic statistics; and 5) the qualitative analysis of translation cases. The code and data are available at https://github.com/NLP2CT/HIL-MT/.

Xin Chen, Junchao Wu, Shu Yang et al.

Instruction Tuning (IT) has been proven to be an effective approach to unlock the powerful capabilities of large language models (LLMs). Recent studies indicate that excessive IT data can degrade LLMs performance, while carefully selecting a small subset of high-quality IT data can significantly enhance their capabilities. Therefore, identifying the most efficient subset data from the IT dataset to effectively develop either specific or general abilities in LLMs has become a critical challenge. To address this, we propose a novel and efficient framework called NAIT. NAIT evaluates the impact of IT data on LLMs performance by analyzing the similarity of neuron activation patterns between the IT dataset and the target domain capability. Specifically, NAIT captures neuron activation patterns from in-domain datasets of target domain capabilities to construct reusable and transferable neuron activation features. It then evaluates and selects optimal samples based on the similarity between candidate samples and the expected activation features of the target capabilities. Experimental results show that training on the 10\% Alpaca-GPT4 IT data subset selected by NAIT consistently outperforms methods that rely on external advanced models or uncertainty-based features across various tasks. Our findings also reveal the transferability of neuron activation features across different capabilities of LLMs. In particular, IT data with more logical reasoning and programmatic features possesses strong general transferability, enabling models to develop stronger capabilities across multiple tasks, while a stable core subset of data is sufficient to consistently activate fundamental model capabilities and universally improve performance across diverse tasks.

Junchao Wu, Runzhe Zhan, Derek F. Wong et al.

Detecting text generated by large language models (LLMs) is of great recent interest. With zero-shot methods like DetectGPT, detection capabilities have reached impressive levels. However, the reliability of existing detectors in real-world applications remains underexplored. In this study, we present a new benchmark, DetectRL, highlighting that even state-of-the-art (SOTA) detection techniques still underperformed in this task. We collected human-written datasets from domains where LLMs are particularly prone to misuse. Using popular LLMs, we generated data that better aligns with real-world applications. Unlike previous studies, we employed heuristic rules to create adversarial LLM-generated text, simulating various prompts usages, human revisions like word substitutions, and writing noises like spelling mistakes. Our development of DetectRL reveals the strengths and limitations of current SOTA detectors. More importantly, we analyzed the potential impact of writing styles, model types, attack methods, the text lengths, and real-world human writing factors on different types of detectors. We believe DetectRL could serve as an effective benchmark for assessing detectors in real-world scenarios, evolving with advanced attack methods, thus providing more stressful evaluation to drive the development of more efficient detectors. Data and code are publicly available at: https://github.com/NLP2CT/DetectRL.

Junchao Wu, Runzhe Zhan, Derek F. Wong et al.

The efficacy of detectors for texts generated by large language models (LLMs) substantially depends on the availability of large-scale training data. However, white-box zero-shot detectors, which require no such data, are limited by the accessibility of the source model of the LLM-generated text. In this paper, we propose a simple yet effective black-box zero-shot detection approach based on the observation that, from the perspective of LLMs, human-written texts typically contain more grammatical errors than LLM-generated texts. This approach involves calculating the Grammar Error Correction Score (GECScore) for the given text to differentiate between human-written and LLM-generated text. Experimental results show that our method outperforms current state-of-the-art (SOTA) zero-shot and supervised methods, achieving an average AUROC of 98.62% across XSum and Writing Prompts dataset. Additionally, our approach demonstrates strong reliability in the wild, exhibiting robust generalization and resistance to paraphrasing attacks. Data and code are available at: https://github.com/NLP2CT/GECScore.

Xinyi Yang, Runzhe Zhan, Derek F. Wong et al.

Investigating bias in large language models (LLMs) is crucial for developing trustworthy AI. While prompt-based through prompt engineering is common, its effectiveness relies on the assumption that models inherently understand biases. Our study systematically analyzed this assumption using the BBQ and StereoSet benchmarks on both open-source models as well as commercial GPT model. Experimental results indicate that prompt-based is often superficial; for instance, the Llama2-7B-Chat model misclassified over 90% of unbiased content as biased, despite achieving high accuracy in identifying bias issues on the BBQ dataset. Additionally, specific evaluation and question settings in bias benchmarks often lead LLMs to choose "evasive answers", disregarding the core of the question and the relevance of the response to the context. Moreover, the apparent success of previous methods may stem from flawed evaluation metrics. Our research highlights a potential "false prosperity" in prompt-base efforts and emphasizes the need to rethink bias metrics to ensure truly trustworthy AI.

Yafu Li, Runzhe Zhan, Haoran Zhang et al.

Recent progress in reasoning models has substantially advanced long-horizon mathematical and scientific problem solving, with several systems now reaching gold-medal-level performance on International Mathematical Olympiad (IMO) and International Physics Olympiad (IPhO) problems. In this paper, we introduce a simple and unified recipe for converting a post-trained reasoning backbone into a rigorous olympiad-level solver. The recipe first uses a reverse-perplexity curriculum for SFT to instill rigorous proof-search and self-checking behaviors, then scales these behaviors through a two-stage RL pipeline that progresses from RL with verifiable rewards to more delicate proof-level RL, and finally boosts solving performance with test-time scaling. Applying this recipe, we train a 30B-A3B backbone with SFT on around 340K sub-8K-token trajectories followed by 200 RL steps. The resulting model, SU-01, supports stable reasoning on difficult problems with trajectories exceeding 100K tokens, while achieving gold-medal-level performance on mathematical and physical olympiad competitions, including IMO 2025/USAMO 2026 and IPhO 2024/2025. It also demonstrates strong generalization of scientific reasoning to domains beyond mathematics and physics.

Xin Chen, Junchao Wu, Shu Yang et al.

Detecting content generated by large language models (LLMs) is crucial for preventing misuse and building trustworthy AI systems. Although existing detection methods perform well, their robustness in out-of-distribution (OOD) scenarios is still lacking. In this paper, we hypothesize that, compared to features used by existing detection methods, the internal representations of LLMs contain more comprehensive and raw features that can more effectively capture and distinguish the statistical pattern differences between LLM-generated texts (LGT) and human-written texts (HWT). We validated this hypothesis across different LLMs and observed significant differences in neural activation patterns when processing these two types of texts. Based on this, we propose RepreGuard, an efficient statistics-based detection method. Specifically, we first employ a surrogate model to collect representation of LGT and HWT, and extract the distinct activation feature that can better identify LGT. We can classify the text by calculating the projection score of the text representations along this feature direction and comparing with a precomputed threshold. Experimental results show that RepreGuard outperforms all baselines with average 94.92% AUROC on both in-distribution (ID) and OOD scenarios, while also demonstrating robust resilience to various text sizes and mainstream attacks. Data and code are publicly available at: https://github.com/NLP2CT/RepreGuard

Yuyi Huang, Runzhe Zhan, Derek F. Wong et al.

Large language models (LLMs) have significantly influenced various industries but suffer from a critical flaw, the potential sensitivity of generating harmful content, which poses severe societal risks. We developed and tested novel attack strategies on popular LLMs to expose their vulnerabilities in generating inappropriate content. These strategies, inspired by psychological phenomena such as the "Priming Effect", "Safe Attention Shift", and "Cognitive Dissonance", effectively attack the models' guarding mechanisms. Our experiments achieved an attack success rate (ASR) of 100% on various open-source models, including Meta's Llama-3.2, Google's Gemma-2, Mistral's Mistral-NeMo, Falcon's Falcon-mamba, Apple's DCLM, Microsoft's Phi3, and Qwen's Qwen2.5, among others. Similarly, for closed-source models such as OpenAI's GPT-4o, Google's Gemini-1.5, and Claude-3.5, we observed an ASR of at least 95% on the AdvBench dataset, which represents the current state-of-the-art. This study underscores the urgent need to reassess the use of generative models in critical applications to mitigate potential adverse societal impacts.

Runzhe Zhan, Xuebo Liu, Derek F. Wong et al.

We release 70 small and discriminative test sets for machine translation (MT) evaluation called variance-aware test sets (VAT), covering 35 translation directions from WMT16 to WMT20 competitions. VAT is automatically created by a novel variance-aware filtering method that filters the indiscriminative test instances of the current MT test sets without any human labor. Experimental results show that VAT outperforms the original WMT test sets in terms of the correlation with human judgement across mainstream language pairs and test sets. Further analysis on the properties of VAT reveals the challenging linguistic features (e.g., translation of low-frequency words and proper nouns) for competitive MT systems, providing guidance for constructing future MT test sets. The test sets and the code for preparing variance-aware MT test sets are freely available at https://github.com/NLP2CT/Variance-Aware-MT-Test-Sets .

Runzhe Zhan, Xuebo Liu, Derek F. Wong et al.

The high-quality translation results produced by machine translation (MT) systems still pose a huge challenge for automatic evaluation. Current MT evaluation pays the same attention to each sentence component, while the questions of real-world examinations (e.g., university examinations) have different difficulties and weightings. In this paper, we propose a novel difficulty-aware MT evaluation metric, expanding the evaluation dimension by taking translation difficulty into consideration. A translation that fails to be predicted by most MT systems will be treated as a difficult one and assigned a large weight in the final score function, and conversely. Experimental results on the WMT19 English-German Metrics shared tasks show that our proposed method outperforms commonly used MT metrics in terms of human correlation. In particular, our proposed method performs well even when all the MT systems are very competitive, which is when most existing metrics fail to distinguish between them. The source code is freely available at https://github.com/NLP2CT/Difficulty-Aware-MT-Evaluation.

Runzhe Zhan, Xuebo Liu, Derek F. Wong et al.

Meta-learning has been sufficiently validated to be beneficial for low-resource neural machine translation (NMT). However, we find that meta-trained NMT fails to improve the translation performance of the domain unseen at the meta-training stage. In this paper, we aim to alleviate this issue by proposing a novel meta-curriculum learning for domain adaptation in NMT. During meta-training, the NMT first learns the similar curricula from each domain to avoid falling into a bad local optimum early, and finally learns the curricula of individualities to improve the model robustness for learning domain-specific knowledge. Experimental results on 10 different low-resource domains show that meta-curriculum learning can improve the translation performance of both familiar and unfamiliar domains. All the codes and data are freely available at https://github.com/NLP2CT/Meta-Curriculum.

Runzhe Zhan, Xinyi Yang, Derek F. Wong et al.

While supervised fine-tuning (SFT) has been a straightforward approach for tailoring the output of foundation large language model (LLM) to specific preferences, concerns have been raised about the depth of this alignment, with some critiques suggesting it is merely "superficial". We critically examine this hypothesis within the scope of cross-lingual generation tasks, proposing that the effectiveness of SFT may be constrained by its reliance on prior tokens to guide cross-lingual generation. Based on this crucial insight, and in response to the challenges posed by the costly and limited availability of non-English data for SFT, we introduce a novel training-free alignment method named PreTTY, which employs minimal task-related prior tokens to bridge the foundation LLM and the SFT LLM, achieving comparable performance without training. Experiments on machine translation and part-of-speech tagging across eight languages demonstrate the efficacy of PreTTY in cross-lingual settings. Remarkably, by initiating the decoding process with only one or two prior tokens, foundation LLMs can achieve performance comparable to their SFT counterparts. This method presents a cost-effective alternative to SFT and advances the democratization of multilingual LLMs.

Jingkun Ma, Runzhe Zhan, Yang Li et al.

A hallmark of advanced artificial intelligence is the capacity to progress from passive visual perception to the strategic modification of visual information to facilitate complex reasoning. This advanced capability, however, remains critically underdeveloped in current Large Multi-modal Models (LMMs). The deficiency is often masked by evaluation metrics that prioritize final-answer accuracy, creating an illusion of competence where genuine reasoning is absent. Using the domain of geometric problem-solving as a precise instrument, we probe this issue through tasks that require constructing visual aids. To this end, we introduce \textbf{VisAidMath}, a challenging benchmark, and our novel Three-Layered Funnel Evaluation Framework. This framework moves beyond simple accuracy (ACCU) to scrutinize the generation of valid visual aids (PVA) and the soundness of subsequent reasoning steps (SPRS). Our extensive experiments on state-of-the-art models, including Doubao-Seed-1.6 and o4, reveal a profound ``Reasoning Illusion''. We observe that high surface-level accuracy conceals a catastrophic failure in the models' ability to produce valid visual aids or to reason from them. Our findings expose a fundamental schism between visual perception and logical deduction in modern LMMs. We host an evaluation platform at CodaBench for testing publicly. Homepage: https://nlp2ct.github.io/VisAidMathHomepage/ Evaluation: https://www.codabench.org/competitions/7634/

Haoyun Xu, Runzhe Zhan, Derek F. Wong et al.

Large Language Models (LLMs) are composed of neurons that exhibit various behaviors and roles, which become increasingly diversified as models scale. Recent studies have revealed that not all neurons are active across different datasets, and this sparsity correlates positively with the task-specific ability, leading to advancements in model pruning and training efficiency. Traditional fine-tuning methods engage all parameters of LLMs, which is computationally expensive and may not be necessary. In contrast, Parameter-Efficient Fine-Tuning (PEFT) approaches aim to minimize the number of trainable parameters, yet they still operate at a relatively macro scale (e.g., layer-level). We introduce Neuron-Level Fine-Tuning (NeFT), a novel approach that refines the granularity of parameter training down to the individual neuron, enabling more precise and computationally efficient model updates. The experimental results show that NeFT not only exceeded the performance of full-parameter fine-tuning and PEFT but also provided insights into the analysis of neurons.

Runzhe Zhan, Zhihong Huang, Xinyi Yang et al.

Recent advancements in large reasoning models (LRMs) have introduced an intermediate "thinking" process prior to generating final answers, improving their reasoning capabilities on complex downstream tasks. However, the potential of LRMs as evaluators for machine translation (MT) quality remains underexplored. We provides the first systematic analysis of LRM-as-a-judge in MT evaluation. We identify key challenges, revealing LRMs require tailored evaluation materials, tend to "overthink" simpler instances and have issues with scoring mechanisms leading to overestimation. To address these, we propose to calibrate LRM thinking by training them on synthetic, human-like thinking trajectories. Our experiments on WMT24 Metrics benchmarks demonstrate that this approach largely reduces thinking budgets by ~35x while concurrently improving evaluation performance across different LRM scales from 7B to 32B (e.g., R1-Distill-Qwen-7B achieves a +8.7 correlation point improvement). These findings highlight the potential of efficiently calibrated LRMs to advance fine-grained automatic MT evaluation.

Yuyi Huang, Runzhe Zhan, Lidia S. Chao et al.

As large language models (LLMs) are increasingly deployed for complex reasoning tasks, Long Chain-of-Thought (Long-CoT) prompting has emerged as a key paradigm for structured inference. Despite early-stage safeguards enabled by alignment techniques such as RLHF, we identify a previously underexplored vulnerability: reasoning trajectories in Long-CoT models can drift from aligned paths, resulting in content that violates safety constraints. We term this phenomenon Path Drift. Through empirical analysis, we uncover three behavioral triggers of Path Drift: (1) first-person commitments that induce goal-driven reasoning that delays refusal signals; (2) ethical evaporation, where surface-level disclaimers bypass alignment checkpoints; (3) condition chain escalation, where layered cues progressively steer models toward unsafe completions. Building on these insights, we introduce a three-stage Path Drift Induction Framework comprising cognitive load amplification, self-role priming, and condition chain hijacking. Each stage independently reduces refusal rates, while their combination further compounds the effect. To mitigate these risks, we propose a path-level defense strategy incorporating role attribution correction and metacognitive reflection (reflective safety cues). Our findings highlight the need for trajectory-level alignment oversight in long-form reasoning beyond token-level alignment.

Runzhe Zhan, Yafu Li, Zhi Wang et al.

Reinforcement learning from verifiable rewards (RLVR) is an emerging paradigm for improving the reasoning ability of large language models. However, standard on-policy training discards rollout experiences after a single update, leading to computational inefficiency and instability. While prior work on RL has highlighted the benefits of reusing past experience, the role of experience characteristics in shaping learning dynamics of large reasoning models remains underexplored. In this paper, we are the first to investigate what makes a reasoning experience valuable and identify rollout correctness and entropy as effective indicators of experience value. Based on these insights, we propose ExGRPO (Experiential Group Relative Policy Optimization), a framework that organizes and prioritizes valuable experiences, and employs a mixed-policy objective to balance exploration with experience exploitation. Experiments on five backbone models (1.5B-8B parameters) show that ExGRPO consistently improves reasoning performance on mathematical/general benchmarks, with an average gain of +3.5/7.6 points over on-policy RLVR. Moreover, ExGRPO stabilizes training on both stronger and weaker models where on-policy methods fail. These results highlight principled experience management as a key ingredient for efficient and scalable RLVR.

Yanming Sun, Runzhe Zhan, Chi Seng Cheang et al.

\textbf{RE}trieval-\textbf{A}ugmented \textbf{L}LM-based \textbf{M}achine \textbf{T}ranslation (REAL-MT) shows promise for knowledge-intensive tasks like idiomatic translation, but its reliability under noisy retrieval contexts remains poorly understood despite this being a common challenge in real-world deployment. To address this gap, we propose a noise synthesis framework and new metrics to evaluate the robustness of REAL-MT systematically. Using this framework, we instantiate REAL-MT with Qwen-series models, including standard LLMs and large reasoning models (LRMs) with enhanced reasoning, and evaluate their performance on idiomatic translation across high-, medium-, and low-resource language pairs under synthesized noise. Our results show that low-resource language pairs, which rely more heavily on retrieved context, degrade more severely under noise than high-resource ones and often produce nonsensical translations. Although LRMs possess enhanced reasoning capabilities, they show no improvement in error correction and are even more susceptible to noise, tending to rationalize incorrect contexts. We find that this stems from an attention shift away from the source idiom to noisy content, while confidence increases despite declining accuracy, indicating poor calibration. To mitigate these issues, we investigate training-free and fine-tuning strategies, which improve robustness at the cost of performance in clean contexts, revealing a fundamental trade-off. Our findings highlight the limitations of current approaches, underscoring the need for self-verifying integration mechanisms.

Zhilin Wang, Zhe Yang, Yun Luo et al.

Enhancing the ability of Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) to interpret sheet music is a crucial step toward building AI musicians. However, current research lacks both evaluation benchmarks and training data for sheet music reasoning. Inspired by mathematics, where simple operations yield infinite verifiable problems, we introduce a novel approach that treats core music theory rules, such as those governing beats and intervals, as programmatic functions to systematically synthesize a vast and diverse corpus of sheet music reasoning problems. This approach allows us to introduce a data synthesis framework that generates verifiable sheet music questions in both textual and visual modalities, leading to the Synthetic Sheet Music Reasoning Benchmark (SSMR-Bench) and a complementary training set. Evaluation results on SSMR-Bench highlight the key role reasoning plays in interpreting sheet music, while also pointing out the ongoing challenges in understanding sheet music in a visual format. By leveraging synthetic data for RLVR, all models show significant improvements on the SSMR-Bench. Additionally, they also demonstrate considerable advancements on previously established human-crafted benchmarks, such as MusicTheoryBench and the music subset of MMMU. Finally, our results show that the enhanced reasoning ability can also facilitate music composition.