Yinghui He

Vedant Shah, Dingli Yu, Kaifeng Lyu et al. · mila, tsinghua

Current LLM training positions mathematical reasoning as a core capability. With publicly available sources fully tapped, there is unmet demand for diverse and challenging math questions. Relying solely on human experts is both time-consuming and costly, while LLM-generated questions often lack the requisite diversity and difficulty. We present a design framework that combines the strengths of LLMs with a human-in-the-loop approach to generate a diverse array of challenging math questions. We leverage LLM metacognition skills [Didolkar et al., 2024] of a strong LLM to extract core "skills" from existing math datasets. These skills serve as the basis for generating novel and difficult questions by prompting the LLM with random pairs of core skills. The use of two different skills within each question makes finding such questions an "out of distribution" task for both LLMs and humans. Our pipeline employs LLMs to iteratively generate and refine questions and solutions through multiturn prompting. Human annotators then verify and further refine the questions, with their efficiency enhanced via further LLM interactions. Applying this pipeline on skills extracted from the MATH dataset [Hendrycks et al., 2021] resulted in MATH$^2$ - a dataset of higher-quality math questions, as evidenced by: (a) Lower performance of all models on MATH$^2$ than on MATH (b) Higher performance on MATH when using MATH$^2$ questions as in-context examples. Although focused on mathematics, our methodology seems applicable to other domains requiring structured reasoning, and potentially as a component of scalable oversight. Also of interest is a striking relationship observed between models' performance on the new dataset: the success rate on MATH$^2$ is the square on MATH, suggesting that successfully solving the question in MATH$^2$ requires a nontrivial combination of two distinct math skills.

Yinghui He, Yufan Wu, Yilin Jia et al.

Theory of Mind (ToM) is the ability to reason about one's own and others' mental states. ToM plays a critical role in the development of intelligence, language understanding, and cognitive processes. While previous work has primarily focused on first and second-order ToM, we explore higher-order ToM, which involves recursive reasoning on others' beliefs. We introduce HI-TOM, a Higher Order Theory of Mind benchmark. Our experimental evaluation using various Large Language Models (LLMs) indicates a decline in performance on higher-order ToM tasks, demonstrating the limitations of current LLMs. We conduct a thorough analysis of different failure cases of LLMs, and share our thoughts on the implications of our findings on the future of NLP.

Yinghui He, Simran Kaur, Adithya Bhaskar et al.

Current post-training methods in verifiable settings fall into two categories. Reinforcement learning (RLVR) relies on binary rewards, which are broadly applicable and powerful, but provide only sparse supervision during training. Distillation provides dense token-level supervision, typically obtained from an external teacher or using high-quality demonstrations. Collecting such supervision can be costly or unavailable. We propose Self-Distillation Zero (SD-Zero), a method that is substantially more training sample-efficient than RL and does not require an external teacher or high-quality demonstrations. SD-Zero trains a single model to play two roles: a Generator, which produces an initial response, and a Reviser, which conditions on that response and its binary reward to produce an improved response. We then perform on-policy self-distillation to distill the reviser into the generator, using the reviser's token distributions conditioned on the generator's response and its reward as supervision. In effect, SD-Zero trains the model to transform binary rewards into dense token-level self-supervision. On math and code reasoning benchmarks with Qwen3-4B-Instruct and Olmo-3-7B-Instruct, SD-Zero improves performance by at least 10% over the base models and outperforms strong baselines, including Rejection Fine-Tuning (RFT), GRPO, and Self-Distillation Fine-Tuning (SDFT), under the same question set and training sample budget. Extensive ablation studies show two novel characteristics of our proposed algorithm: (a) token-level self-localization, where the reviser can identify the key tokens that need to be revised in the generator's response based on reward, and (b) iterative self-evolution, where the improving ability to revise answers can be distilled back into generation performance with regular teacher synchronization.

Xi Ye, Fangcong Yin, Yinghui He et al.

Existing benchmarks for evaluating long-context language models (LCLMs) primarily focus on long-context recall, requiring models to produce short responses based on a few critical snippets while processing thousands of irrelevant tokens. We introduce LongProc (Long Procedural Generation), a new benchmark that requires both the integration of highly dispersed information and long-form generation. LongProc consists of six diverse procedural generation tasks, such as extracting structured information from HTML pages into a TSV format and executing complex search procedures to create travel plans. These tasks challenge LCLMs by testing their ability to follow detailed procedural instructions, synthesize and reason over dispersed information, and generate structured, long-form outputs (up to 8K tokens). Furthermore, as these tasks adhere to deterministic procedures and yield structured outputs, they enable reliable rule-based evaluation. We evaluated 23 LCLMs, including instruction-tuned models and recent reasoning models, on LongProc at three difficulty levels, with the maximum number of output tokens set at 500, 2K, and 8K. Notably, while all tested models claim a context window size above 32K tokens, open-weight models typically falter on 2K-token tasks, and closed-source models like GPT-4o show significant degradation on 8K-token tasks. Reasoning models achieve stronger overall performance in long-form generation, benefiting from long CoT training. Further analysis reveals that LCLMs struggle to maintain long-range coherence in long-form generations. These findings highlight critical limitations in current LCLMs and suggest substantial room for improvement. Data and code available at: https://princeton-pli.github.io/LongProc.

Jiahao Qiu, Yinghui He, Xinzhe Juan et al.

The rise of LLM-driven AI characters raises safety concerns, particularly for vulnerable human users with psychological disorders. To address these risks, we propose EmoAgent, a multi-agent AI framework designed to evaluate and mitigate mental health hazards in human-AI interactions. EmoAgent comprises two components: EmoEval simulates virtual users, including those portraying mentally vulnerable individuals, to assess mental health changes before and after interactions with AI characters. It uses clinically proven psychological and psychiatric assessment tools (PHQ-9, PDI, PANSS) to evaluate mental risks induced by LLM. EmoGuard serves as an intermediary, monitoring users' mental status, predicting potential harm, and providing corrective feedback to mitigate risks. Experiments conducted in popular character-based chatbots show that emotionally engaging dialogues can lead to psychological deterioration in vulnerable users, with mental state deterioration in more than 34.4% of the simulations. EmoGuard significantly reduces these deterioration rates, underscoring its role in ensuring safer AI-human interactions. Our code is available at: https://github.com/1akaman/EmoAgent

Jiarui Liu, Lechen Zhang, Yongjin Yang et al.

Supervised fine-tuning (SFT) is widely used to inject new knowledge into language models, but it often degrades pretrained capabilities such as reasoning and general-domain performance. We argue this forgetting arises because fine-tuning targets from humans or external systems diverge from the model's autoregressive distribution, forcing the optimizer to imitate low-probability token sequences. To address this problem, we propose MixSD, a simple external-teacher-free method for distribution-aligned knowledge injection. Instead of training on fixed targets, MixSD constructs supervision dynamically by mixing tokens from two conditionals of the base model itself: an expert conditional that observes the injected fact in context, and a naive conditional that reflects the model's original prior. The resulting supervision sequences preserve the factual learning signal while remaining substantially closer to the base model's distribution. We evaluate MixSD on two synthetic corpora that we construct to study factual recall and arithmetic function acquisition in a controlled setting, together with established benchmarks for open-domain factual question answering and knowledge editing. Across multiple model scales and settings, MixSD consistently achieves a better memorization-retention trade-off compared to SFT and on-policy self distillation baselines, retaining up to 100% of the base model's held-out capability while maintaining near-perfect training accuracy, whereas standard SFT retains as little as 1%. We further show that MixSD produces substantially lower-NLL supervision targets under the base model and reduces harmful movement along Fisher-sensitive parameter directions. These results suggest that aligning supervision with the model's native generation distribution is a simple and effective principle for knowledge injection that mitigates catastrophic forgetting.

Yinghui He, Abhishek Panigrahi, Yong Lin et al.

Language models often show little to no improvement (i.e., "saturation") when trained via vanilla supervised fine-tuning (SFT) on data similar to what they saw in their training set (e.g., MATH). We introduce a new fine-tuning strategy, STAT, to train such a student model by using the metacognition ability of a stronger large language model (LLM) as the teacher. The teacher uses the task dataset to create a list of skills needed for the task, and then labels each data point with its required skills (Didolkar et al., 2024). By monitoring the student's answers, the teacher creates a Missing-Skill-Profile for the student, tracking how often they failed to apply each skill in their responses. We use this idea to build a modified training set in one of two ways. In STAT-Sel, the teacher uses an existing set of training examples but adaptively reweights them according to the Missing-Skill-Profile. In STAT-Syn, the teacher synthesizes additional examples involving missing skills. Across extensive experiments on Llama and Qwen models, our methods yield improvements of up to 7.5% on MATH, whereas SFT provides only limited gains. Furthermore, STAT enhances performance on out-of-distribution benchmarks (e.g., AIME24/25, AMC23, etc.) by an average of 4.6%. Crucially, we find that STAT is complementary to RL via GRPO (Shao et al., 2024): after the model is improved using STAT to address skill gaps, GRPO continues to add further gains. We conclude that skill-targeted adaptive training should broadly improve current training pipelines. Our code is available at: https://github.com/princeton-pli/STAT.

Fei Deng, Yinghui He, Chuntong Chu et al.

Human Activity Recognition (HAR) in smart homes is critical for health monitoring and assistive living. While vision-based systems are common, they face privacy concerns and environmental limitations (e.g., occlusion). In this work, we present MobiDiary, a framework that generates natural language descriptions of daily activities directly from heterogeneous physical signals (specifically IMU and Wi-Fi). Unlike conventional approaches that restrict outputs to pre-defined labels, MobiDiary produces expressive, human-readable summaries. To bridge the semantic gap between continuous, noisy physical signals and discrete linguistic descriptions, we propose a unified sensor encoder. Instead of relying on modality-specific engineering, we exploit the shared inductive biases of motion-induced signals--where both inertial and wireless data reflect underlying kinematic dynamics. Specifically, our encoder utilizes a patch-based mechanism to capture local temporal correlations and integrates heterogeneous placement embedding to unify spatial contexts across different sensors. These unified signal tokens are then fed into a Transformer-based decoder, which employs an autoregressive mechanism to generate coherent action descriptions word-by-word. We comprehensively evaluate our approach on multiple public benchmarks (XRF V2, UWash, and WiFiTAD). Experimental results demonstrate that MobiDiary effectively generalizes across modalities, achieving state-of-the-art performance on captioning metrics (e.g., BLEU@4, CIDEr, RMC) and outperforming specialized baselines in continuous action understanding.

Yinghui He, Abhishek Panigrahi, Yong Lin et al.

In-context learning (ICL) allows a language model to improve its problem-solving capability when provided with suitable information in context. Since the choice of in-context information can be determined based on the problem itself, in-context learning is analogous to human learning from teachers in a classroom. Recent works (Didolkar et al., 2024a; 2024b) show that ICL performance can be improved by leveraging a frontier large language model's (LLM) ability to predict required skills to solve a problem, popularly referred to as an LLM's metacognition, and using the recommended skills to construct necessary in-context examples. While this skill-based strategy boosts ICL performance in larger models, its gains on small language models (SLMs) have been minimal, highlighting a performance gap in ICL capabilities. We investigate this gap and show that skill-based prompting can hurt SLM performance on easy questions by introducing unnecessary information, akin to cognitive overload. To address this, we introduce AdaptMI, an adaptive approach to selecting skill-based in-context Math Instructions for SLMs. Inspired by cognitive load theory from human pedagogy, our method only introduces skill-based examples when the model performs poorly. We further propose AdaptMI+, which adds examples targeted to the specific skills missing from the model's responses. On 5-shot evaluations across popular math benchmarks and five SLMs (1B--7B; Qwen, Llama), AdaptMI+ improves accuracy by up to 6% over naive skill-based strategies.

Qiqi Xiao, Ziqi Ye, Yinghui He et al.

To promote the practicality of deep learning-based localization, existing studies aim to address the issue of scenario dependence through meta-learning. However, these studies primarily focus on variations in environmental layouts while overlooking the impact of changes in device configurations, such as bandwidth, the number of access points (APs), and the number of antennas used. Unlike environmental changes, variations in device configurations affect the dimensionality of channel state information (CSI), thereby compromising neural network usability. To address this issue, we propose Meta-SimGNN, a novel WiFi localization system that integrates graph neural networks with meta-learning to improve localization generalization and robustness. First, we introduce a fine-grained CSI graph construction scheme, where each AP is treated as a graph node, allowing for adaptability to changes in the number of APs. To structure the features of each node, we propose an amplitude-phase fusion method and a feature extraction method. The former utilizes both amplitude and phase to construct CSI images, enhancing data reliability, while the latter extracts dimension-consistent features to address variations in bandwidth and the number of antennas. Second, a similarity-guided meta-learning strategy is developed to enhance adaptability in diverse scenarios. The initial model parameters for the fine-tuning stage are determined by comparing the similarity between the new scenario and historical scenarios, facilitating rapid adaptation of the model to the new localization scenario. Extensive experimental results over commodity WiFi devices in different scenarios show that Meta-SimGNN outperforms the baseline methods in terms of localization generalization and accuracy.

Jiatong Yu, Yinghui He, Anirudh Goyal et al.

Machine unlearning seeks to selectively remove the "influence" of specific training data on a model's outputs. The ideal goal is Retrain Equivalence--behavior identical to a model trained from scratch on only the retained data. This goal was formulated for models trained on i.i.d. data batches, but modern pipelines often involve multi-stage training, with each stage having a distinct data distribution and objective. Examples include LLM fine-tuning for alignment, reasoning ability, etc. Our study shows via theory and experiments that this shift to multi-stage training introduces a fundamental barrier for machine unlearning. The theory indicates that the outcome of local unlearning--methods that only use gradients computed on the forget set--is path-dependent. That is, a model's behavior during unlearning is influenced by the order of its training stages during learning, making it impossible for path-oblivious algorithms to universally achieve Retrain Equivalence. We empirically demonstrate the same phenomenon in LLM post-training across Llama and Qwen models (1B to 14B) with gradient ascent, NPO, and SimNPO local unlearning algorithms. Models fine-tuned via different orderings of identical training stages diverge in behavior during unlearning, with the degradation in GSM8K accuracy after unlearning varying by over 20% across paths. We also observe that some learning paths consistently produce models that unlearn slowly. During unlearning, whether the probability mass gets squeezed into paraphrasing or alternative concepts is also path-dependent. These results consistently show that Retrain Equivalence is an ill-posed target for local unlearning algorithms, so long as the target models are trained in stages. In situations where access to models' training histories is hard, the current work calls for rethinking the definition and desiderata of machine unlearning.

Xin Li, Jingzhi Hu, Yinghui He et al.

Wi-Fi-based human activity recognition (HAR) provides substantial convenience and has emerged as a thriving research field, yet the coarse spatial resolution inherent to Wi-Fi significantly hinders its ability to distinguish multiple subjects. By exploiting the near-field domination effect, establishing a dedicated sensing link for each subject through their personal Wi-Fi device offers a promising solution for multi-person HAR under native traffic. However, due to the subject-specific characteristics and irregular patterns of near-field signals, HAR neural network models require fine-tuning (FT) for cross-domain adaptation, which becomes particularly challenging with certain categories unavailable. In this paper, we propose WiAnchor, a novel training framework for efficient cross-domain adaptation in the presence of incomplete activity categories. This framework processes Wi-Fi signals embedded with irregular time information in three steps: during pre-training, we enlarge inter-class feature margins to enhance the separability of activities; in the FT stage, we innovate an anchor matching mechanism for cross-domain adaptation, filtering subject-specific interference informed by incomplete activity categories, rather than attempting to extract complete features from them; finally, the recognition of input samples is further improved based on their feature-level similarity with anchors. We construct a comprehensive dataset to thoroughly evaluate WiAnchor, achieving over 90% cross-domain accuracy with absent activity categories.

Jianhao Ma, Rui Ray Chen, Yinghui He et al.

In this paper, we study the problem of sparse mean estimation under adversarial corruptions, where the goal is to estimate the $k$-sparse mean of a heavy-tailed distribution from samples contaminated by adversarial noise. Existing methods face two key limitations: they require prior knowledge of the sparsity level $k$ and scale poorly to high-dimensional settings. We propose a simple and scalable estimator that addresses both challenges. Specifically, it learns the $k$-sparse mean without knowing $k$ in advance and operates in near-linear time and memory with respect to the ambient dimension. Under a moderate signal-to-noise ratio, our method achieves the optimal statistical rate, matching the information-theoretic lower bound. Extensive simulations corroborate our theoretical guarantees. At the heart of our approach is an incremental learning phenomenon: we show that a basic subgradient method applied to a nonconvex two-layer formulation with an $\ell_1$-loss can incrementally learn the $k$ nonzero components of the true mean while suppressing the rest. More broadly, our work is the first to reveal the incremental learning phenomenon of the subgradient method in the presence of heavy-tailed distributions and adversarial corruption.

Jinke Ren, Yinghui He, Dingzhu Wen et al.

In cellular federated edge learning (FEEL), multiple edge devices holding local data jointly train a neural network by communicating learning updates with an access point without exchanging their data samples. With very limited communication resources, it is beneficial to schedule the most informative local learning updates. In this paper, a novel scheduling policy is proposed to exploit both diversity in multiuser channels and diversity in the "importance" of the edge devices' learning updates. First, a new probabilistic scheduling framework is developed to yield unbiased update aggregation in FEEL. The importance of a local learning update is measured by its gradient divergence. If one edge device is scheduled in each communication round, the scheduling policy is derived in closed form to achieve the optimal trade-off between channel quality and update importance. The probabilistic scheduling framework is then extended to allow scheduling multiple edge devices in each communication round. Numerical results obtained using popular models and learning datasets demonstrate that the proposed scheduling policy can achieve faster model convergence and higher learning accuracy than conventional scheduling policies that only exploit a single type of diversity.