Ning Miao

Xuan Yang, Hao Xu, Tingfeng Hui et al.

Despite great advances in tool-use capabilities of large language models (LLMs), existing evaluation benchmarks struggle to fully align with real-world scenarios. Such benchmarks mostly rely on simulated idealized user assumptions and lacks experience-oriented evaluation. These limitations fail to account for the ambiguity, uncooperative behaviors, and shifting intentions characteristic of real-world users. To fill this gap, we propose RUT-Bench, a dedicated benchmark designed to assess LLMs under diverse Real-world User Tool calling scenarios. RUT-Bench supports high-fidelity simulations covering both ideal rational patterns and heterogeneous non-ideal behaviors across single-turn and multi-turn dialogues. We conduct comprehensive evaluations on 19 widely adopted open-source and proprietary LLMs using our benchmark. Experimental results reveal that no tested LLMs achieve an overall success rate above 40%, and nearly all of them experience noticeable performance drops when facing more complicated non-ideal user inputs. Our code and data is available at https://github.com/Miaow-Lab/RUT-Bench.

Ning Miao, Tom Rainforth, Emile Mathieu et al. · microsoft-research, oxford

We introduce InstaAug, a method for automatically learning input-specific augmentations from data. Previous methods for learning augmentations have typically assumed independence between the original input and the transformation applied to that input. This can be highly restrictive, as the invariances we hope our augmentation will capture are themselves often highly input dependent. InstaAug instead introduces a learnable invariance module that maps from inputs to tailored transformation parameters, allowing local invariances to be captured. This can be simultaneously trained alongside the downstream model in a fully end-to-end manner, or separately learned for a pre-trained model. We empirically demonstrate that InstaAug learns meaningful input-dependent augmentations for a wide range of transformation classes, which in turn provides better performance on both supervised and self-supervised tasks.

Ning Miao, Yee Whye Teh, Tom Rainforth

The recent progress in large language models (LLMs), especially the invention of chain-of-thought prompting, has made it possible to automatically answer questions by stepwise reasoning. However, when faced with more complicated problems that require non-linear thinking, even the strongest LLMs make mistakes. To address this, we explore whether LLMs are able to recognize errors in their own step-by-step reasoning, without resorting to external resources. To this end, we propose SelfCheck, a general-purpose zero-shot verification schema for recognizing such errors. We then use the results of these checks to improve question-answering performance by conducting weighted voting on multiple solutions to the question. We test SelfCheck on three datasets (GSM8K, MathQA, and MATH) and find that it successfully recognizes errors and, in turn, increases final answer accuracies.

Xuan Yang, Jiayu Liu, Yuhang Lai et al.

Large Language Models (LLMs) have achieved strong complex reasoning capabilities through Chain-of-Thought (CoT) reasoning. However, their reasoning patterns remain too complicated to analyze. While Sparse Autoencoders (SAEs) have emerged as a powerful tool for interpretability, existing approaches predominantly operate at the token level, creating a granularity mismatch when capturing more critical step-level information, such as reasoning direction and semantic transitions. In this work, we propose step-level sparse autoencoder (SSAE), which serves as an analytical tool to disentangle different aspects of LLMs' reasoning steps into sparse features. Specifically, by precisely controlling the sparsity of a step feature conditioned on its context, we form an information bottleneck in step reconstruction, which splits incremental information from background information and disentangles it into several sparsely activated dimensions. Experiments on multiple base models and reasoning tasks show the effectiveness of the extracted features. By linear probing, we can easily predict surface-level information, such as generation length and first token distribution, as well as more complicated properties, such as the correctness and logicality of the step. These observations indicate that LLMs should already at least partly know about these properties during generation, which provides the foundation for the self-verification ability of LLMs. The code is available at https://github.com/Miaow-Lab/SSAE

Jialin Liu, Ning Miao, Chongzhou Fang et al.

The Electrocardiogram (ECG) measures the electrical cardiac activity generated by the heart to detect abnormal heartbeat and heart attack. However, the irregular occurrence of the abnormalities demands continuous monitoring of heartbeats. Machine learning techniques are leveraged to automate the task to reduce labor work needed during monitoring. In recent years, many companies have launched products with ECG monitoring and irregular heartbeat alert. Among all classification algorithms, the time series-based algorithm dynamic time warping (DTW) is widely adopted to undertake the ECG classification task. Though progress has been achieved, the DTW-based ECG classification also brings a new attacking vector of leaking the patients' diagnosis results. This paper shows that the ECG input samples' labels can be stolen via a side-channel attack, Flush+Reload. In particular, we first identify the vulnerability of DTW for ECG classification, i.e., the correlation between warping path choice and prediction results. Then we implement an attack that leverages Flush+Reload to monitor the warping path selection with known ECG data and then build a predictor for constructing the relation between warping path selection and labels of input ECG samples. Based on experiments, we find that the Flush+Reload-based inference leakage can achieve an 84.0\% attacking success rate to identify the labels of the two samples in DTW.

Tianle Wang, Zhongyuan Wu, Shenghao Jin et al.

Reinforcement learning with verifiable rewards (RLVR) has become a central component of large language model (LLM) post-training. Unlike supervised fine-tuning (SFT), RLVR lets an LLM generate multiple candidate solutions and reinforces those that lead to a verifiably correct final answer. However, in practice, RLVR often requires thousands of training steps to reach strong performance, incurring substantial computation largely attributed to prolonged exploration. In this work, we make a surprising observation: during RLVR, LLMs evolve in a strongly linear manner. Specifically, both model weights and model output log-probabilities exhibit strong linear correlations with RL training steps. This suggests that RLVR predominantly amplifies trends that emerge early in training, rather than continuously discovering new behaviors throughout the entire optimization trajectory. Motivated by this linearity, we investigate whether future model states can be predicted from intermediate checkpoints via extrapolation, avoiding continued expensive training. We show that Weight Extrapolation produces models with performance comparable to standard RL training while requiring significantly less computation. Moreover, Logits Extrapolation consistently outperforms continued RL training on mathematics and code benchmarks by extrapolating beyond the step range where RL training remains stable. Our code is available at https://github.com/Miaow-Lab/RLVR-Linearity

Tingfeng Hui, Hao Xu, Pengyu Zhu et al.

Large language models (LLMs) deployed in real-world agentic applications must be capable of replanning and adapting when mid-task disruptions invalidate their prior decisions. Existing dynamic benchmarks primarily measure whether LLMs can detect temporal changes in a timely manner, leaving the complementary challenge of adaptive replanning under spatio-temporal dynamics largely unexplored. We introduce STT-Arena (Spatio-Temporal Tool-Use Arena), a benchmark of 227 high-quality interactive tasks spanning nine spatio-temporal conflict types and four solvability levels. Each task is grounded in a realistic, executable environment equipped with injected spatio-temporal triggers that can abruptly invalidate an ongoing plan, forcing the model to detect the state shift and construct a revised execution strategy. Extensive evaluation of frontier LLMs reveals that even the SOTA proprietary models, including Claude-4.6-Opus, achieves less than 40\% overall accuracies, highlighting the fundamental difficulty of spatio-temporal dynamic reasoning. Systematic analysis of failure trajectories uncovers three recurring error modes of existing models: Stale-State Execution, Misdiagnosis of Dynamic Triggers, and Missing Post-Adaptation Verification. Guided by these findings, we propose an iterative trajectory refinement technique that eliminates these failure patterns from training data, and combine it with online RL to produce STT-Agent-4B which outperforms frontier LLMs on STT-Arena.

Ning Miao, Yuxuan Song, Hao Zhou et al.

It has been a common approach to pre-train a language model on a large corpus and fine-tune it on task-specific data. In practice, we observe that fine-tuning a pre-trained model on a small dataset may lead to over- and/or under-estimation problem. In this paper, we propose MC-Tailor, a novel method to alleviate the above issue in text generation tasks by truncating and transferring the probability mass from over-estimated regions to under-estimated ones. Experiments on a variety of text generation datasets show that MC-Tailor consistently and significantly outperforms the fine-tuning approach. Our code is available at this url.

Wenxian Shi, Hao Zhou, Ning Miao et al.

Deep generative models are commonly used for generating images and text. Interpretability of these models is one important pursuit, other than the generation quality. Variational auto-encoder (VAE) with Gaussian distribution as prior has been successfully applied in text generation, but it is hard to interpret the meaning of the latent variable. To enhance the controllability and interpretability, one can replace the Gaussian prior with a mixture of Gaussian distributions (GM-VAE), whose mixture components could be related to hidden semantic aspects of data. In this paper, we generalize the practice and introduce DEM-VAE, a class of models for text generation using VAEs with a mixture distribution of exponential family. Unfortunately, a standard variational training algorithm fails due to the mode-collapse problem. We theoretically identify the root cause of the problem and propose an effective algorithm to train DEM-VAE. Our method penalizes the training with an extra dispersion term to induce a well-structured latent space. Experimental results show that our approach does obtain a meaningful space, and it outperforms strong baselines in text generation benchmarks. The code is available at https://github.com/wenxianxian/demvae.

Ning Miao, Hao Zhou, Lili Mou et al.

In real-world applications of natural language generation, there are often constraints on the target sentences in addition to fluency and naturalness requirements. Existing language generation techniques are usually based on recurrent neural networks (RNNs). However, it is non-trivial to impose constraints on RNNs while maintaining generation quality, since RNNs generate sentences sequentially (or with beam search) from the first word to the last. In this paper, we propose CGMH, a novel approach using Metropolis-Hastings sampling for constrained sentence generation. CGMH allows complicated constraints such as the occurrence of multiple keywords in the target sentences, which cannot be handled in traditional RNN-based approaches. Moreover, CGMH works in the inference stage, and does not require parallel corpora for training. We evaluate our method on a variety of tasks, including keywords-to-sentence generation, unsupervised sentence paraphrasing, and unsupervised sentence error correction. CGMH achieves high performance compared with previous supervised methods for sentence generation. Our code is released at https://github.com/NingMiao/CGMH

Yuhang Lai, Jiazhan Feng, Yee Whye Teh et al.

Large Language Models (LLMs) demonstrate strong capabilities for solving scientific and mathematical problems, yet they struggle to produce valid, challenging, and novel problems - an essential component for advancing LLM training and enabling autonomous scientific research. Existing problem generation approaches either depend on expensive human expert involvement or adopt naive self-play paradigms, which frequently yield invalid problems due to reward hacking. This work introduces VHG, a verifier-enhanced hard problem generation framework built upon three-party self-play. By integrating an independent verifier into the conventional setter-solver duality, our design constrains the setter's reward to be jointly determined by problem validity (evaluated by the verifier) and difficulty (assessed by the solver). We instantiate two verifier variants: a Hard symbolic verifier and a Soft LLM-based verifier, with evaluations conducted on indefinite integral tasks and general mathematical reasoning tasks. Experimental results show that VHG substantially outperforms all baseline methods by a clear margin.

Deepro Choudhury, Sinead Williamson, Adam Goliński et al. · oxford

We propose a general-purpose approach for improving the ability of Large Language Models (LLMs) to intelligently and adaptively gather information from a user or other external source using the framework of sequential Bayesian experimental design (BED). This enables LLMs to act as effective multi-turn conversational agents and interactively interface with external environments. Our approach, which we call BED-LLM (Bayesian Experimental Design with Large Language Models), is based on iteratively choosing questions or queries that maximize the expected information gain (EIG) about the task of interest given the responses gathered previously. We show how this EIG can be formulated (and then estimated) in a principled way using a probabilistic model derived from the LLM's predictive distributions and provide detailed insights into key decisions in its construction and updating procedure. We find that BED-LLM achieves substantial gains in performance across a wide range of tests based on the 20 questions game and using the LLM to actively infer user preferences, compared to direct prompting of the LLM and other adaptive design strategies.

Wei Shao, Zequan Liang, Ruoyu Zhang et al.

Biometric authentication using physiological signals offers a promising path toward secure and user-friendly access control in wearable devices. While electrocardiogram (ECG) signals have shown high discriminability, their intrusive sensing requirements and discontinuous acquisition limit practicality. Photoplethysmography (PPG), on the other hand, enables continuous, non-intrusive authentication with seamless integration into wrist-worn wearable devices. However, most prior work relies on high-frequency PPG (e.g., 75 - 500 Hz) and complex deep models, which incur significant energy and computational overhead, impeding deployment in power-constrained real-world systems. In this paper, we present the first real-world implementation and evaluation of a continuous authentication system on a smartwatch, We-Be Band, using low-frequency (25 Hz) multi-channel PPG signals. Our method employs a Bi-LSTM with attention mechanism to extract identity-specific features from short (4 s) windows of 4-channel PPG. Through extensive evaluations on both public datasets (PTTPPG) and our We-Be Dataset (26 subjects), we demonstrate strong classification performance with an average test accuracy of 88.11%, macro F1-score of 0.88, False Acceptance Rate (FAR) of 0.48%, False Rejection Rate (FRR) of 11.77%, and Equal Error Rate (EER) of 2.76%. Our 25 Hz system reduces sensor power consumption by 53% compared to 512 Hz and 19% compared to 128 Hz setups without compromising performance. We find that sampling at 25 Hz preserves authentication accuracy, whereas performance drops sharply at 20 Hz while offering only trivial additional power savings, underscoring 25 Hz as the practical lower bound. Additionally, we find that models trained exclusively on resting data fail under motion, while activity-diverse training improves robustness across physiological states.

Jiayu Liu, Wei Dai, Zhenya Huang et al.

Despite the strong reasoning ability of large language models~(LLMs), they are prone to errors and hallucinations. As a result, how to check their outputs effectively and efficiently has become a critical problem in their applications. Existing checking methods heavily rely on external resources, such as trained verifiers (e.g., process/outcome reward models) or elaborate prompts, which lead to high computational overhead and are only applicable to specific domains. In this paper, we investigate whether the internal behaviors of LLMs have already implied the credibility of their reasoning paths. Specifically, we find that the rank of the correlation matrix between the input problem and the output reasoning path is a robust indicator of reasoning correctness. Different from other correctness indicators for LLMs, the calculation of the correlation matrix only relies on the LLM itself, which avoids the hassle of training a separate model or designing complicated prompts. Based on it, we design a simple, plug-and-play Self-Indicator method to reweight candidate reasoning paths, which achieves significant performance improvements than other voting and verification methods with very few computational overhead. Our experiments across multiple LLMs of varying scales and model families have further shown the effectiveness of Self-Indicator. It achieves over 75% accuracy in distinguishing correct reasoning paths from incorrect ones, and, in turn, improves the accuracies on three reasoning benchmarks by more than 8%.

Qiyuan Liu, Hao Xu, Xuhong Chen et al.

Reward models (RMs) play a critical role in enhancing the reasoning performance of LLMs. For example, they can provide training signals to finetune LLMs during reinforcement learning (RL) and help select the best answer from multiple candidates during inference. In this paper, we provide a systematic introduction to RMs, along with a comprehensive survey of their applications in LLM reasoning. We first review fundamental concepts of RMs, including their architectures, training methodologies, and evaluation techniques. Then, we explore their key applications: (1) guiding generation and selecting optimal outputs during LLM inference, (2) facilitating data synthesis and iterative self-improvement for LLMs, and (3) providing training signals in RL-based finetuning. Finally, we discuss critical open questions regarding the selection, generalization, evaluation, and enhancement of RMs, based on existing research and our own empirical findings. Our analysis aims to provide actionable insights for the effective deployment and advancement of RMs for LLM reasoning.

Jiayu Liu, Zhenya Huang, Anya Sims et al.

The current paradigm for reasoning in large language models (LLMs) involves models "thinking out loud" via a sequence of tokens, known as chain-of-thought (CoT). This approach, while effective, has several significant drawbacks. Firstly, inference requires autoregressive generation of often thousands of CoT tokens, which is slow and computationally expensive. Secondly, it constrains reasoning to the discrete space of tokens, creating an information bottleneck across reasoning steps. Thirdly, it fundamentally entangles reasoning with token generation, forcing LLMs to "think while speaking," which causes potentially short-sighted reasoning. In light of these limitations, we re-imagine reasoning in LLMs and present a new paradigm: MARCOS. In our approach, rather than autoregressively generating tokens, we model reasoning as a hidden Markov chain of continuous, high-dimensional "thoughts". Each reasoning step involves a transition of the internal thoughts, where explicit reasoning steps (which may consist of hundreds of tokens) serve as observable variables, which are windows to peek into the implicit thoughts. Since this latent process is incompatible with the standard supervised learning, we further propose a two-phase variational training scheme. Our experiments on three benchmarks demonstrate that MARCOS outperforms existing continuous reasoning methods and, for the first time, achieves performance comparable to token-based CoT, even surpassing it by 4.7% on GSM8K with up to 15.7x speedup in inference. Beyond this, MARCOS offers additional advantages, such as step-level instead of token-level control over randomness, opening significant opportunities for reinforcement learning and reasoning in LLMs.

Ning Miao, Emile Mathieu, N. Siddharth et al.

We explain why directly changing the prior can be a surprisingly ineffective mechanism for incorporating inductive biases into VAEs, and introduce a simple and effective alternative approach: Intermediary Latent Space VAEs(InteL-VAEs). InteL-VAEs use an intermediary set of latent variables to control the stochasticity of the encoding process, before mapping these in turn to the latent representation using a parametric function that encapsulates our desired inductive bias(es). This allows us to impose properties like sparsity or clustering on learned representations, and incorporate human knowledge into the generative model. Whereas changing the prior only indirectly encourages behavior through regularizing the encoder, InteL-VAEs are able to directly enforce desired characteristics. Moreover, they bypass the computation and encoder design issues caused by non-Gaussian priors, while allowing for additional flexibility through training of the parametric mapping function. We show that these advantages, in turn, lead to both better generative models and better representations being learned.

Huangzhao Zhang, Hao Zhou, Ning Miao et al.

Efficiently building an adversarial attacker for natural language processing (NLP) tasks is a real challenge. Firstly, as the sentence space is discrete, it is difficult to make small perturbations along the direction of gradients. Secondly, the fluency of the generated examples cannot be guaranteed. In this paper, we propose MHA, which addresses both problems by performing Metropolis-Hastings sampling, whose proposal is designed with the guidance of gradients. Experiments on IMDB and SNLI show that our proposed MHA outperforms the baseline model on attacking capability. Adversarial training with MAH also leads to better robustness and performance.

Yuxuan Song, Ning Miao, Hao Zhou et al.

Auto-regressive sequence generative models trained by Maximum Likelihood Estimation suffer the exposure bias problem in practical finite sample scenarios. The crux is that the number of training samples for Maximum Likelihood Estimation is usually limited and the input data distributions are different at training and inference stages. Many method shave been proposed to solve the above problem (Yu et al., 2017; Lu et al., 2018), which relies on sampling from the non-stationary model distribution and suffers from high variance or biased estimations. In this paper, we proposeψ-MLE, a new training scheme for auto-regressive sequence generative models, which is effective and stable when operating at large sample space encountered in text generation. We derive our algorithm from a new perspective of self-augmentation and introduce bias correction with density ratio estimation. Extensive experimental results on synthetic data and real-world text generation tasks demonstrate that our method stably outperforms Maximum Likelihood Estimation and other state-of-the-art sequence generative models in terms of both quality and diversity.

Ning Miao, Hao Zhou, Chengqi Zhao et al.

Neural models for text generation require a softmax layer with proper token embeddings during the decoding phase. Most existing approaches adopt single point embedding for each token. However, a word may have multiple senses according to different context, some of which might be distinct. In this paper, we propose KerBS, a novel approach for learning better embeddings for text generation. KerBS embodies two advantages: (a) it employs a Bayesian composition of embeddings for words with multiple senses; (b) it is adaptive to semantic variances of words and robust to rare sentence context by imposing learned kernels to capture the closeness of words (senses) in the embedding space. Empirical studies show that KerBS significantly boosts the performance of several text generation tasks.