Chunyang Xiao

Xin Liu, Runsong Zhao, Xinyu Liu et al.

Extended chain-of-thought (CoT) traces improve LLM reasoning but incur substantial computational and memory costs. While existing CoT compression methods mitigate this by condensing thought steps into compact representations via memory tokens and retaining only these representations at inference time, the loss of fine-grained information makes subsequent steps more error-prone. To alleviate this, we propose \textbf{HybridThinker}, where in addition to preserved these representations, thought steps are also temporarily retained to provide fine-grained details. However, we observe that naively keeping thought steps accessible to subsequent steps \emph{during training} lets the model bypass memory tokens by retrieving information directly from these steps, leaving the model's ability to compress and retrieve information through memory tokens insufficiently trained. We therefore introduce a hybrid training scheme, in which only some thought steps are directly accessible through attention to subsequent steps, while the other thought steps are masked, forcing the model to use memory tokens for compression and retrieval. Across 4 reasoning benchmarks, HybridThinker matches the uncompressed baseline, advancing the state of the art in CoT compression by 5.8 points on average accuracy with similar inference time. Ablation studies confirm that both temporary thought-step retention and the hybrid training scheme contribute to these gains.

Xinyu Liu, Yan Ding, Kaikai An et al. · pku

While state-of-the-art NLP models have demonstrated excellent performance for aspect based sentiment analysis (ABSA), substantial evidence has been presented on their lack of robustness. This is especially manifested as significant degradation in performance when faced with out-of-distribution data. Recent solutions that rely on counterfactually augmented datasets show promising results, but they are inherently limited because of the lack of access to explicit causal structure. In this paper, we present an alternative approach that relies on non-counterfactual data augmentation. Our proposal instead relies on using noisy, cost-efficient data augmentations that preserve semantics associated with the target aspect. Our approach then relies on modelling invariances between different versions of the data to improve robustness. A comprehensive suite of experiments shows that our proposal significantly improves upon strong pre-trained baselines on both standard and robustness-specific datasets. Our approach further establishes a new state-of-the-art on the ABSA robustness benchmark and transfers well across domains.

Sae Young Moon, Myeongjun Erik Jang, Haoyan Luo et al.

Topic modeling has extensive applications in text mining and data analysis across various industrial sectors. Although the concept of granularity holds significant value for business applications by providing deeper insights, the capability of topic modeling methods to produce granular topics has not been thoroughly explored. In this context, this paper introduces a framework called TIDE, which primarily provides a novel granular topic modeling method based on large language models (LLMs) as a core feature, along with other useful functionalities for business applications, such as summarizing long documents, topic parenting, and distillation. Through extensive experiments on a variety of public and real-world business datasets, we demonstrate that TIDE's topic modeling approach outperforms modern topic modeling methods, and our auxiliary components provide valuable support for dealing with industrial business scenarios. The TIDE framework is currently undergoing the process of being open sourced.

Runsong Zhao, Xin Liu, Xinyu Liu et al.

Using special tokens (e.g., gist, memory, or compressed tokens) to compress context information is a common practice for large language models (LLMs). However, existing approaches often neglect that position encodings inherently induce local inductive biases in models, causing the compression process to ignore holistic contextual dependencies. We propose \textbf{Enhanced Position Layout (EPL)}, a simple yet effective method that improves the context compression capability of LLMs by only adjusting position IDs, the numerical identifiers that specify token positions. EPL minimizes the distance between context tokens and their corresponding special tokens and at the same time maintains the sequence order in position IDs between context tokens, special tokens, and the subsequent tokens. Integrating EPL into our best performing context compression model results in a 1.9 ROUGE-1 F1 improvement on out-of-domain question answering datasets on average. When extended to multimodal scenarios, EPL leads to an average accuracy gain of 2.6 points for vision compression LLMs.

Pengcheng Huang, Yongyu Mu, Yuzhang Wu et al.

Imposing constraints on machine translation systems presents a challenging issue because these systems are not trained to make use of constraints in generating adequate, fluent translations. In this paper, we leverage the capabilities of large language models (LLMs) for constrained translation, given that LLMs can easily adapt to this task by taking translation instructions and constraints as prompts. However, LLMs cannot always guarantee the adequacy of translation, and, in some cases, ignore the given constraints. This is in part because LLMs might be overly confident in their predictions, overriding the influence of the constraints. To overcome this overiding behaviour, we propose to add a revision process that encourages LLMs to correct the outputs by prompting them about the constraints that have not yet been met. We evaluate our approach on four constrained translation tasks, encompassing both lexical and structural constraints in multiple constraint domains. Experiments show 15\% improvement in constraint-based translation accuracy over standard LLMs and the approach also significantly outperforms neural machine translation (NMT) state-of-the-art methods.

Thomas Walshe, Sae Young Moon, Chunyang Xiao et al.

Acquiring labelled training data remains a costly task in real world machine learning projects to meet quantity and quality requirements. Recently Large Language Models (LLMs), notably GPT-4, have shown great promises in labelling data with high accuracy. However, privacy and cost concerns prevent the ubiquitous use of GPT-4. In this work, we explore effectively leveraging open-source models for automatic labelling. We identify integrating label schema as a promising technology but found that naively using the label description for classification leads to poor performance on high cardinality tasks. To address this, we propose Retrieval Augmented Classification (RAC) for which LLM performs inferences for one label at a time using corresponding label schema; we start with the most related label and iterates until a label is chosen by the LLM. We show that our method, which dynamically integrates label description, leads to performance improvements in labelling tasks. We further show that by focusing only on the most promising labels, RAC can trade off between label quality and coverage - a property we leverage to automatically label our internal datasets.

Xinyu Liu, Xin Liu, Bo Jin et al.

While Chain-of-thought (CoT) reasoning enables LLMs to solve challenging reasoning problems, as KV cache grows linearly with the number of generated tokens, CoT reasoning faces scaling issues in terms of speed and memory usage. In this work, we propose MemoSight (Memory-Foresight-based reasoning), a unified framework that integrates both context compression and multi-token prediction to mitigate the efficiency issues while maintaining CoT reasoning performance. Our framework adopts the same minimalist design for both context compression and multi-token prediction via special tokens and their corresponding position layout tailored to each token type. Comprehensive experiments on four reasoning benchmarks demonstrate that MemoSight reduces the KV cache footprint by up to 66% and accelerates inference by 1.56x, while outperforming existing CoT compression methods.

Xinyu Liu, Kechen Jiao, Chunyang Xiao et al.

The convergence of reinforcement learning and imitation learning has positioned Reverse KL (RKL) as a promising paradigm for on-policy LLM distillation, aiming to unify exploration with teacher supervision. However, we identify a critical limitation: when the student and teacher distributions diverge significantly, standard RKL often fails to yield meaningful improvement due to uninformative negative feedback. To address this inefficiency, we propose Teacher-Guided Policy Optimization (TGPO), an on-policy algorithm that incorporates dense directional guidance by leveraging teacher predictions conditioned on the student's rollout. Because TGPO remains on-policy, the algorithm integrates seamlessly with existing RLVR frameworks without requiring additional data annotation. Experiments on complex reasoning benchmarks demonstrate that TGPO significantly outperforms standard baselines and is robust to different teachers.

Xin Liu, Runsong Zhao, Pengcheng Huang et al.

Context compression presents a promising approach for accelerating large language model (LLM) inference by compressing long contexts into compact representations. Current context compression methods predominantly rely on autoencoding tasks to train context-agnostic compression tokens to compress contextual semantics. While autoencoding tasks enable compression tokens to acquire compression capabilities, compression via autoencoding tasks creates a fundamental mismatch: the models are optimized for reconstruction that diverge from actual downstream tasks, thereby weakening the features more beneficial for real-world usage. We propose Semantic-Anchor Compression (SAC), a novel method that shifts from autoencoding task based compression to an architecture that is equipped with this compression capability \textit{a priori}. Instead of training models to compress contexts through autoencoding tasks, SAC directly selects so-called anchor tokens from the original context and aggregates contextual information into their key-value (KV) representations. By deriving representations directly from the contextual tokens, SAC eliminates the need for autoencoding training. To ensure compression performance while directly leveraging anchor tokens, SAC incorporates two key designs: (1) anchor embeddings that enable the compressor to identify critical tokens, and (2) bidirectional attention modification that allows anchor tokens to capture information from the entire context. Experimental results demonstrate that SAC consistently outperforms existing context compression methods across various compression ratios. On out-of-distribution evaluation using MRQA, SAC achieves 1 EM improvement at 5x compression over strong baselines, with increasing advantages at higher compression ratios.

Chunyang Xiao, Pranava Madhyastha

In this paper we present a controlled study on the linearized IRM framework (IRMv1) introduced in Arjovsky et al. (2020). We show that IRMv1 (and its variants) framework can be potentially unstable under small changes to the optimal regressor. This can, notably, lead to worse generalisation to new environments, even compared with ERM which converges simply to the global minimum for all training environments mixed up all together. We also highlight the isseus of scaling in the the IRMv1 setup. These observations highlight the importance of rigorous evaluation and importance of unit-testing for measuring progress towards IRM.

Chunyang Xiao, Christoph Teichmann, Konstantine Arkoudas

While sequence-to-sequence (seq2seq) models achieve state-of-the-art performance in many natural language processing tasks, they can be too slow for real-time applications. One performance bottleneck is predicting the most likely next token over a large vocabulary; methods to circumvent this bottleneck are a current research topic. We focus specifically on using seq2seq models for semantic parsing, where we observe that grammars often exist which specify valid formal representations of utterance semantics. By developing a generic approach for restricting the predictions of a seq2seq model to grammatically permissible continuations, we arrive at a widely applicable technique for speeding up semantic parsing. The technique leads to a 74% speed-up on an in-house dataset with a large vocabulary, compared to the same neural model without grammatical restrictions.

Chunyang Xiao, Marc Dymetman, Claire Gardent

Seq2seq models based on Recurrent Neural Networks (RNNs) have recently received a lot of attention in the domain of Semantic Parsing for Question Answering. While in principle they can be trained directly on pairs (natural language utterances, logical forms), their performance is limited by the amount of available data. To alleviate this problem, we propose to exploit various sources of prior knowledge: the well-formedness of the logical forms is modeled by a weighted context-free grammar; the likelihood that certain entities present in the input utterance are also present in the logical form is modeled by weighted finite-state automata. The grammar and automata are combined together through an efficient intersection algorithm to form a soft guide ("background") to the RNN. We test our method on an extension of the Overnight dataset and show that it not only strongly improves over an RNN baseline, but also outperforms non-RNN models based on rich sets of hand-crafted features.

Marc Dymetman, Chunyang Xiao

We introduce LL-RNNs (Log-Linear RNNs), an extension of Recurrent Neural Networks that replaces the softmax output layer by a log-linear output layer, of which the softmax is a special case. This conceptually simple move has two main advantages. First, it allows the learner to combat training data sparsity by allowing it to model words (or more generally, output symbols) as complex combinations of attributes without requiring that each combination is directly observed in the training data (as the softmax does). Second, it permits the inclusion of flexible prior knowledge in the form of a priori specified modular features, where the neural network component learns to dynamically control the weights of a log-linear distribution exploiting these features. We conduct experiments in the domain of language modelling of French, that exploit morphological prior knowledge and show an important decrease in perplexity relative to a baseline RNN. We provide other motivating iillustrations, and finally argue that the log-linear and the neural-network components contribute complementary strengths to the LL-RNN: the LL aspect allows the model to incorporate rich prior knowledge, while the NN aspect, according to the "representation learning" paradigm, allows the model to discover novel combination of characteristics.

Chunyang Xiao

In an online decision problem, one makes decisions often with a pool of decision sequence called experts but without knowledge of the future. After each step, one pays a cost based on the decision and observed rate. One reasonal goal would be to perform as well as the best expert in the pool. The modern and well-known way to attain this goal is the algorithm of exponential weighting. However, recently, another algorithm called follow the perturbed leader is developed and achieved about the same performance. In our work, we first show the properties shared in common by the two algorithms which explain the similarities on the performance. Next we will show that for a specific perturbation, the two algorithms are identical. Finally, we show with some examples that follow-the-leader style algorithms extend naturally to a large class of structured online problems for which the exponential algorithms are inefficient.