Zhengyan Shi

Minseon Kim, Lucas Caccia, Zhengyan Shi et al.

User prompts to large language models (LLMs) are often ambiguous or under-specified, and subtle contextual cues shaped by user intentions, prior knowledge, and risk factors strongly influence what constitutes an appropriate response. Misinterpreting intent or risks may lead to unsafe outputs, while overly cautious interpretations can cause unnecessary refusal of benign requests. In this paper, we question the conventional framework in which LLMs generate immediate responses to requests without considering broader contextual factors. User requests are situated within broader contexts such as intentions, knowledge, and prior experience, which strongly influence what constitutes an appropriate answer. We propose a framework that extracts and leverages such contextual information from the user prompt itself. Specifically, a reinforcement learning based context generator, designed in an autoencoder-like fashion, is trained to infer contextual signals grounded in the prompt and use them to guide response generation. This approach is particularly important for safety tasks, where ambiguous requests may bypass safeguards while benign but confusing requests can trigger unnecessary refusals. Experiments show that our method reduces harmful responses by an average of 5.6% on the SafetyInstruct dataset across multiple foundation models and improves the harmonic mean of attack success rate and compliance on benign prompts by 6.2% on XSTest and WildJailbreak. These results demonstrate the effectiveness of context extraction for safer and more reliable LLM inferences.

Wanru Zhao, Lucas Caccia, Zhengyan Shi et al.

Curriculum learning is a class of training strategies that organizes the data being exposed to a model by difficulty, gradually from simpler to more complex examples. This research explores a reverse curriculum generation approach that recursively decomposes complex datasets into simpler, more learnable components. We propose a teacher-student framework where the teacher is equipped with the ability to reason step-by-step, which is used to recursively generate easier versions of examples, enabling the student model to progressively master difficult tasks. We propose a novel scoring system to measure data difficulty based on its structural complexity and conceptual depth, allowing curriculum construction over decomposed data. Experiments on math datasets (MATH and AIME) and code generation datasets demonstrate that models trained with curricula generated by our approach exhibit superior performance compared to standard training on original datasets.

Hyunji Lee, Minseon Kim, Chinmay Singh et al.

As coding agents are increasingly deployed in large codebases, the need to automatically design challenging, codebase-level evaluation is central. We propose Gistify, a task where a coding LLM must create a single, minimal, self-contained file that can reproduce a specific functionality of a codebase. The coding LLM is given full access to a codebase along with a specific entrypoint (e.g., a python command), and the generated file must replicate the output of the same command ran under the full codebase, while containing only the essential components necessary to execute the provided command. Success on Gistify requires both structural understanding of the codebase, accurate modeling of its execution flow as well as the ability to produce potentially large code patches. Our findings show that current state-of-the-art models struggle to reliably solve Gistify tasks, especially ones with long executions traces.

Kai Jiang, Zhengyan Shi, Dell Zhang et al.

Class Incremental Learning (CIL) aims to continuously learn new categories while retaining the knowledge of old ones. Pre-trained models (PTMs) show promising capabilities in CIL. However, existing approaches that apply lightweight fine-tuning to backbones still induce parameter drift, thereby compromising the generalization capability of pre-trained models. Parameter drift can be conceptualized as a form of noise that obscures critical patterns learned for previous tasks. However, recent researches have shown that noise is not always harmful. For example, the large number of visual patterns learned from pre-training can be easily abused by a single task, and introducing appropriate noise can suppress some low-correlation features, thus leaving a margin for future tasks. To this end, we propose learning beneficial noise for CIL guided by information theory and propose Mixture of Noise (Min), aiming to mitigate the degradation of backbone generalization from adapting new tasks. Specifically, task-specific noise is learned from high-dimension features of new tasks. Then, a set of weights is adjusted dynamically for optimal mixture of different task noise. Finally, Min embeds the beneficial noise into the intermediate features to mask the response of inefficient patterns. Extensive experiments on six benchmark datasets demonstrate that Min achieves state-of-the-art performance in most incremental settings, with particularly outstanding results in 50-steps incremental settings. This shows the significant potential for beneficial noise in continual learning. Code is available at https://github.com/ASCIIJK/MiN-NeurIPS2025.

Jianrong Ding, Jianyuan Zhong, Zhengyan Shi et al.

Large Language Models (LLMs) are increasingly used to brainstorm and evaluate research ideas, yet assessing such judgments is fundamentally difficult because the true impact of a new idea may take years to emerge. We address this challenge by using the impact forecasting of human-authored manuscripts as a verifiable proxy task. In a prospective forecasting study, we find that frontier LLMs fail to reliably distinguish high-impact papers from ordinary publications, suggesting that static text-based judging is insufficient for scientific evaluation. To address this limitation, we propose $\textbf{FAME}$ ($\underline{\text{F}}$orecasting $\underline{\text{A}}$cademic Impact via Continuous-Time $\underline{\text{M}}$anifold $\underline{\text{E}}$volution), a spatiotemporal framework for modeling the dynamic trajectories of scientific topics. FAME projects papers into a dynamic latent space informed by textual features and a verified knowledge-flow graph, learning geometric constraints that align impactful manuscripts with the forward momentum of their fields. Experiments on 3,200 arXiv papers across three fast-evolving subfields show that FAME consistently and substantially outperforms state-of-the-art LLM evaluators in prospective multidimensional impact forecasting. Furthermore, integrating FAME's dynamic geometric signals into LLMs significantly improves their forecasting performance. These results support manuscript impact forecasting as a useful, measurable proxy benchmark and position FAME as a strong, trajectory-aware foundation for automated scientific evaluation.

Zhengyan Shi, Adam X. Yang, Bin Wu et al.

Instruction tuning plays a crucial role in shaping the outputs of language models (LMs) to desired styles. In this work, we propose a simple yet effective method, Instruction Modelling (IM), which trains LMs by applying a loss function to the instruction and prompt part rather than solely to the output part. Through experiments across 21 diverse benchmarks, we show that, in many scenarios, IM can effectively improve the LM performance on both NLP tasks (e.g., MMLU, TruthfulQA, and HumanEval) and open-ended generation benchmarks (e.g., MT-Bench and AlpacaEval). Remarkably, in the most advantageous case, IM boosts model performance on AlpacaEval 1.0 by over 100%. We identify two key factors influencing the effectiveness of IM: (1) The ratio between instruction length and output length in the training data; and (2) The number of training examples. We observe that IM is especially beneficial when trained on datasets with lengthy instructions paired with brief outputs, or under the Superficial Alignment Hypothesis (SAH) where a small amount of training examples are used for instruction tuning. Further analysis substantiates our hypothesis that our improvement can be attributed to reduced overfitting to instruction tuning datasets. It is worth noting that we are not proposing \ours as a replacement for current fine-tuning processes. Instead, our work aims to provide practical guidance for instruction tuning LMs, especially in low-resource scenarios.

Adam X. Yang, Maxime Robeyns, Thomas Coste et al.

To ensure that large language model (LLM) responses are helpful and non-toxic, a reward model trained on human preference data is usually used. LLM responses with high rewards are then selected through best-of-$n$ (BoN) sampling or the LLM is further optimized to produce responses with high rewards through reinforcement learning from human feedback (RLHF). However, these processes are susceptible to reward overoptimization or `hacking', where responses receive high rewards due to imperfections in the reward model rather than true preference, particularly as prompts or responses deviate from the training data. To address these challenges, we propose to train a Bayesian reward model, which signals higher uncertainty further from the training data distribution. We trained Bayesian reward models using Laplace approximation on LoRA weights, and found that the resulting uncertainty estimates can effectively mitigate reward overoptimization in BoN sampling.

Zhengyan Shi, Sander Land, Acyr Locatelli et al.

Direct Alignment Algorithms (DAAs), such as Direct Preference Optimisation (DPO) and Identity Preference Optimisation (IPO), have emerged as alternatives to online Reinforcement Learning from Human Feedback (RLHF) algorithms such as Proximal Policy Optimisation (PPO) for aligning language models to human preferences, without the need for explicit reward modelling. These methods generally aim to increase the likelihood of generating better (preferred) completions while discouraging worse (non-preferred) ones, while staying close to the original model's behaviour. In this work, we explore the relationship between completion likelihood and model performance in state-of-the-art DAAs, and identify a critical issue of likelihood over-optimisation. Contrary to expectations, we find that higher likelihood of better completions and larger margins between better and worse completion likelihoods do not necessarily lead to better performance, and may even degrade it. Our analysis reveals that while higher likelihood correlates with better memorisation of factual knowledge patterns, a slightly lower completion likelihood tends to improve output diversity, thus leading to better generalisation to unseen scenarios. Moreover, we identify two key indicators that signal when over-optimised output diversity begins to harm performance: Decreasing Entropy over Top-k Tokens and Diminishing Top-k Probability Mass. Our experimental results validate that these indicators are reliable signs of declining performance under different regularisations, helping prevent over-optimisation and improve alignment with human preferences.

Atharv Sonwane, Isadora White, Hyunji Lee et al.

High quality bugs are key to training the next generation of language model based software engineering (SWE) agents. We introduce a novel method for synthetic generation of difficult and diverse bugs. Our method instructs SWE Agents to introduce a feature into the codebase whereby they may unintentionally break tests, resulting in bugs. Prior approaches often induce an out-of-distribution effect by generating bugs intentionally (e.g. by introducing local perturbation to existing code), which does not reflect realistic development processes. We perform qualitative analysis to demonstrate that our approach for generating bugs more closely reflects the patterns found in human-authored edits. Through extensive experiments, we demonstrate that our bugs provide more efficient training data for supervised fine-tuning, outperforming other bug datasets by 2% with half the training data (1.2k vs. 3k bugs). We train on our newly generated bugs in addition to existing bug datasets to get FrogBoss a state-of-the-art 32B parameter model on SWE-bench Verified with a pass@1 of 54.6% and FrogMini a state-of-the-art 14B model on SWE-bench Verified with a pass@1 of 45.3% on SWE-bench Verified averaged over three seeds.

Chenyi Zhou, Zhengyan Shi, Yuan Yao et al.

Recent advancements in large language models (LLMs) have highlighted their potential across a variety of tasks, but their performance still heavily relies on the design of effective prompts. Existing methods for automatic prompt optimization face two challenges: lack of diversity, limiting the exploration of valuable and innovative directions and semantic drift, where optimizations for one task can degrade performance in others. To address these issues, we propose Residual Optimization Tree (RiOT), a novel framework for automatic prompt optimization. RiOT iteratively refines prompts through text gradients, generating multiple semantically diverse candidates at each step, and selects the best prompt using perplexity. Additionally, RiOT incorporates the text residual connection to mitigate semantic drift by selectively retaining beneficial content across optimization iterations. A tree structure efficiently manages the optimization process, ensuring scalability and flexibility. Extensive experiments across five benchmarks, covering commonsense, mathematical, logical, temporal, and semantic reasoning, demonstrate that RiOT outperforms both previous prompt optimization methods and manual prompting.

Yuchen Zhu, Daniel Augusto de Souza, Zhengyan Shi et al.

We address the problem of reward hacking, where maximising a proxy reward does not necessarily increase the true reward. This is a key concern for Large Language Models (LLMs), as they are often fine-tuned on human preferences that may not accurately reflect a true objective. Existing work uses various tricks such as regularisation, tweaks to the reward model, and reward hacking detectors, to limit the influence that such proxy preferences have on a model. Luckily, in many contexts such as medicine, education, and law, a sparse amount of expert data is often available. In these cases, it is often unclear whether the addition of proxy data can improve policy learning. We outline a set of sufficient conditions on proxy feedback that, if satisfied, indicate that proxy data can provably improve the sample complexity of learning the ground truth policy. These conditions can inform the data collection process for specific tasks. The result implies a parameterisation for LLMs that achieves this improved sample complexity. We detail how one can adapt existing architectures to yield this improved sample complexity.

Boyuan Chen, Sitong Fang, Jiaming Ji et al.

As intelligence increases, so does its shadow. AI deception, in which systems induce false beliefs to secure self-beneficial outcomes, has evolved from a speculative concern to an empirically demonstrated risk across language models, AI agents, and emerging frontier systems. This project provides a comprehensive and up-to-date overview of the AI deception field, covering its core concepts, methodologies, genesis, and potential mitigations. First, we identify a formal definition of AI deception, grounded in signaling theory from studies of animal deception. We then review existing empirical studies and associated risks, highlighting deception as a sociotechnical safety challenge. We organize the landscape of AI deception research as a deception cycle, consisting of two key components: deception emergence and deception treatment. Deception emergence reveals the mechanisms underlying AI deception: systems with sufficient capability and incentive potential inevitably engage in deceptive behaviors when triggered by external conditions. Deception treatment, in turn, focuses on detecting and addressing such behaviors. On deception emergence, we analyze incentive foundations across three hierarchical levels and identify three essential capability preconditions required for deception. We further examine contextual triggers, including supervision gaps, distributional shifts, and environmental pressures. On deception treatment, we conclude detection methods covering benchmarks and evaluation protocols in static and interactive settings. Building on the three core factors of deception emergence, we outline potential mitigation strategies and propose auditing approaches that integrate technical, community, and governance efforts to address sociotechnical challenges and future AI risks. To support ongoing work in this area, we release a living resource at www.deceptionsurvey.com.

Zhengyan Shi

Language models (LMs) have demonstrated remarkable capabilities in NLP, yet adapting them efficiently and robustly to specific tasks remains challenging. As their scale and complexity grow, fine-tuning LMs on labelled data often underutilizes available unlabelled data, leads to overfitting on small task-specific sets, and imposes significant computational costs. These limitations hamper their application to the open-ended landscape of real-world language tasks. This thesis proposes a series of methods to better adapt LMs to downstream applications. First, we explore strategies for extracting task-relevant knowledge from unlabelled data, introducing a novel continued pre-training technique that outperforms state-of-the-art semi-supervised approaches. Next, we present a parameter-efficient fine-tuning method that substantially reduces memory and compute costs while maintaining competitive performance. We also introduce improved supervised fine-tuning methods that enable LMs to better follow instructions, especially when labelled data is scarce, enhancing their performance across a range of NLP tasks, including open-ended generation. Finally, we develop new evaluation methods and benchmarks, such as multi-hop spatial reasoning tasks, to assess LM capabilities and adaptation more comprehensively. Through extensive empirical studies across diverse NLP tasks, our results demonstrate that these approaches substantially improve LM robustness, efficiency, and generalization, making them more adaptable to a broad range of applications. These advances mark a significant step towards more robust and efficient LMs, bringing us closer to the goal of artificial general intelligence.

Bin Wu, Zhengyan Shi, Hossein A. Rahmani et al.

Utilizing user profiles to personalize Large Language Models (LLMs) has been shown to enhance the performance on a wide range of tasks. However, the precise role of user profiles and their effect mechanism on LLMs remains unclear. This study first confirms that the effectiveness of user profiles is primarily due to personalization information rather than semantic information. Furthermore, we investigate how user profiles affect the personalization of LLMs. Within the user profile, we reveal that it is the historical personalized response produced or approved by users that plays a pivotal role in personalizing LLMs. This discovery unlocks the potential of LLMs to incorporate a greater number of user profiles within the constraints of limited input length. As for the position of user profiles, we observe that user profiles integrated into different positions of the input context do not contribute equally to personalization. Instead, where the user profile that is closer to the beginning affects more on the personalization of LLMs. Our findings reveal the role of user profiles for the personalization of LLMs, and showcase how incorporating user profiles impacts performance providing insight to leverage user profiles effectively.