arXiv subjects

Chia-Hsuan Lee

Publications and source records attributed to Chia-Hsuan Lee.

17 recordsLinked to original sources

arXiv2026

Know When to Stop: Segment-Level Credit Assignment for Reducing Overthinking

Reasoning language models frequently overthink: generating extended chains of behaviors such as hedging, approach abandonment, and self contradiction that consume tokens without improving answers. We show that these behaviors are not merely a consequence of length; even when controlling for response length, incorrect traces exhibit higher rates of unproductive self-reflection than correct ones. Addressing this requires identifying where self-reflection helps vs hurts, but obtaining these step-level annotations is costly. We observe that intermediate answer commitments within reasoning traces can provide a cheap proxy: by comparing each final answer candidate in the trace to the ground truth, we can determine whether subsequent reflection is productive without any additional supervision. Building on this insight, we propose DASH (Drift Aware advantage SHaping), which assigns segment-level credit based on whether each reasoning segment leads toward or away from correctness. On competition-level math benchmarks, DASH achieves the highest accuracy where overthinking is prevalent (Average Accuracy: 59.45% vs. 58.1% Dr.GRPO vs. 56.95% GRPO) while reducing overthinking behaviors and achieving more productive self-correction than baselines.

cs.CL↗

arXiv2026

Structured Thoughts For Improved Reasoning And Context Pruning

Large language models (LLMs) excel at generating long chains of thought, but long reasoning traces are often verbose and memory-inefficient. In this work, we introduce Structured Thoughts, a framework that organizes reasoning into alternating and blocks: captures exploratory scratch work, while contains the distilled conclusion of that step. We construct a dataset of structured thoughts by segmenting reasoning traces into blocks and prompting an LLM to summarize each step into its corresponding . Fine-tuning pretrained foundation models on this reformatted data produces models that adopt the structured reasoning style, leading to performance gains of up to 8.08\% on reasoning benchmarks compared to standard SFT. The explicit structure also enables context pruning: after each / pair, the can be pruned, allowing the model to retain conclusions without keeping the full scratch work in the context. A proof-of-concept pruning implementation achieves an average of 85\% memory / context savings with an 8.67\% performance drop across mathematical tasks.

cs.CL↗

arXiv2026

SEAD: Competence-Aware On-Policy Distillation via Entropy-Guided Supervision

On-policy distillation (OPD) has a property absent in offline distillation and RL: teacher supervision quality depends on student competence. Incoherent rollouts yield noisy gradients; already-mastered tokens yield redundant ones. This creates waste at three scales (tokens, training phases, and prompts) yet existing methods supervise uniformly. We introduce SEAD, which uses entropy as a unified probe of this competence-dependent degradation at three scales: (1) joint teacher-student entropy partitions tokens into zones receiving tailored divergences or zero gradient (approx. 50% skipped); (2) a cosine schedule anneals from forward to reverse KL as competence grows; (3) a competence-gated curriculum introduces prompts easy-to-hard. These components are symbiotically necessary: token selection requires coherent rollouts (curriculum), annealing requires monotonic improvement (also curriculum). On OLMo-3 (7B to 32B), SEAD achieves +4.8 avg accuracy over vanilla OPD across six math benchmarks, with ablations confirming super-additive interactions.

cs.CL↗

arXiv2026

Critique-Guided Distillation for Robust Reasoning via Refinement

Supervised fine-tuning with expert demonstrations often produces models that imitate outputs without internalizing the reasoning processes needed for robust generalization. While critique-based approaches show promise, training models to generate critiques directly, such as Critique Fine-Tuning (CFT), can lead to output-format drift and degradation of general capabilities. We propose Critique-Guided Distillation (CGD), a training framework that decouples critique consumption from critique generation. During fine-tuning, the student is trained to refine flawed responses conditioned on teacher critiques. CGD treats critiques as a \textit{training-time-only} supervision signal, encouraging internalization of error-aware reasoning: critiques guide learning but are absent at inference. Controlled ablations confirm that these reasoning gains are directly driven by the specificity and relevance of the teacher's feedback. Across five model families, CGD consistently outperforms CFT and standard distillation on mathematical reasoning benchmarks, yielding 7\% average improvements and gains of up to +15.0\% on AMC23 and +12.2\% on MATH-500. On challenging competition problems such as AIME24 and AIME25, CGD achieves substantially higher Pass@1 and stronger performance at low Pass@k, indicating improved reasoning quality per sample. Importantly, CGD preserves general instruction-following capabilities where CFT degrades significantly ($-$21.3\% on IFEval). These results position CGD as a practical and compute-efficient intermediate training paradigm for reasoning-centric tasks without introducing architectural inference-time overhead.

cs.CL↗

arXiv2026

Decomposing the Delta: What Do Models Actually Learn from Preference Pairs?

Preference optimization methods such as DPO and KTO are widely used for aligning language models, yet little is understood about what properties of preference data drive downstream reasoning gains. We ask: what aspects of a preference pair improve a reasoning model's performance on general reasoning tasks? We investigate two distinct notions of quality delta in preference data: generator-level delta, arising from the differences in capability between models that generate chosen and rejected reasoning traces, and sample-level delta, arising from differences in judged quality differences within an individual preference pair. To study generator-level delta, we vary the generator's scale and model family, and to study sample-level delta, we employ an LLM-as-a-judge to rate the quality of generated traces along multiple reasoning-quality dimensions. We find that increasing generator-level delta steadily improves performance on out-of-domain reasoning tasks and filtering data by sample-level delta can enable more data-efficient training. Our results suggest a twofold recipe for improving reasoning performance through preference optimization: maximize generator-level delta when constructing preference pairs and exploit sample-level delta to select the most informative training examples.

cs.CL↗

arXiv2026

DIAL-SUMMER: A Structured Evaluation Framework of Hierarchical Errors in Dialogue Summaries

Dialogues are a predominant mode of communication for humans, and it is immensely helpful to have automatically generated summaries of them (e.g., to revise key points discussed in a meeting, to review conversations between customer agents and product users). Prior works on dialogue summary evaluation largely ignore the complexities specific to this task: (i) shift in structure, from multiple speakers discussing information in a scattered fashion across several turns, to a summary's sentences, and (ii) shift in narration viewpoint, from speakers' first/second-person narration, standardized third-person narration in the summary. In this work, we introduce our framework DIALSUMMER to address the above. We propose DIAL-SUMMER's taxonomy of errors to comprehensively evaluate dialogue summaries at two hierarchical levels: DIALOGUE-LEVEL that focuses on the broader speakers/turns, and WITHIN-TURN-LEVEL that focuses on the information talked about inside a turn. We then present DIAL-SUMMER's dataset composed of dialogue summaries manually annotated with our taxonomy's fine-grained errors. We conduct empirical analyses of these annotated errors, and observe interesting trends (e.g., turns occurring in middle of the dialogue are the most frequently missed in the summary, extrinsic hallucinations largely occur at the end of the summary). We also conduct experiments on LLM-Judges' capability at detecting these errors, through which we demonstrate the challenging nature of our dataset, the robustness of our taxonomy, and the need for future work in this field to enhance LLMs' performance in the same. Code and inference dataset coming soon.

cs.CL↗

arXiv2024

CorrectionLM: Self-Corrections with SLM for Dialogue State Tracking

Large language models (LLMs) have demonstrated self-improvement capabilities via feedback and refinement, but current small language models (SLMs) have had limited success in this area. Existing correction approaches often rely on distilling knowledge from LLMs, which imposes significant computation demands. In this work, we introduce CORRECTIONLM, a novel correction framework that enables SLMs to self-correct using in-context exemplars without LLM involvement. Applied to two dialogue state tracking (DST) tasks in low-resource settings, CORRECTIONLM achieves results similar to a state-of-the-art LLM at a small fraction of the computation costs.

cs.CL↗

arXiv2024

OrchestraLLM: Efficient Orchestration of Language Models for Dialogue State Tracking

Large language models (LLMs) have revolutionized the landscape of Natural Language Processing systems, but are computationally expensive. To reduce the cost without sacrificing performance, previous studies have explored various approaches to harness the potential of Small Language Models (SLMs) as cost-effective alternatives to their larger counterparts. Driven by findings that SLMs and LLMs exhibit complementary strengths in a structured knowledge extraction task, this work presents a novel SLM/LLM routing framework designed to improve computational efficiency and enhance task performance. First, exemplar pools are created to represent the types of contexts where each LM provides a more reliable answer, leveraging a sentence embedding fine-tuned so that context similarity is close to dialogue state similarity. Then, during inference, the k-nearest exemplars to the testing instance are retrieved, and the instance is routed according to majority vote. In dialogue state tracking tasks, the proposed routing framework enhances performance substantially compared to relying solely on LLMs, while reducing the computational costs by over 50%.

cs.CL↗

arXiv2024

Does Collaborative Human-LM Dialogue Generation Help Information Extraction from Human Dialogues?

The capabilities of pretrained language models have opened opportunities to explore new application areas, but applications involving human-human interaction are limited by the fact that most data is protected from public release for privacy reasons. Problem-solving human dialogues in real applications can be much more complex than existing Wizard-of-Oz collections, preventing successful domain transfer. To support information extraction (IE) for a private call center dataset, we introduce a human-in-the-loop dialogue generation framework capable of synthesizing realistic dialogues. In IE experiments with auto insurance call center dialogues, we observe 25\% relative improvement in $F_1$ after augmenting a small set of real human conversations with synthetic data. We release code and our synthetic dataset to illustrate the complexity of real-world call center conversations and encourage development of complex dialogue datasets that are more representative of natural data.

cs.CL↗

arXiv2022

In-Context Learning for Few-Shot Dialogue State Tracking

Collecting and annotating task-oriented dialogues is time-consuming and costly; thus, zero and few shot learning could greatly benefit dialogue state tracking (DST). In this work, we propose an in-context learning (ICL) framework for zero-shot and few-shot learning DST, where a large pre-trained language model (LM) takes a test instance and a few exemplars as input, and directly decodes the dialogue state without any parameter updates. To better leverage a tabular domain description in the LM prompt, we reformulate DST into a text-to-SQL problem. We also propose a novel approach to retrieve annotated dialogues as exemplars. Empirical results on MultiWOZ show that our method IC-DST substantially outperforms previous fine-tuned state-of-the-art models in few-shot settings. In addition, we test IC-DST in zero-shot settings, in which the model only takes a fixed task instruction as input, finding that it outperforms previous zero-shot methods by a large margin.

cs.CL↗

arXiv2022

MIA 2022 Shared Task: Evaluating Cross-lingual Open-Retrieval Question Answering for 16 Diverse Languages

We present the results of the Workshop on Multilingual Information Access (MIA) 2022 Shared Task, evaluating cross-lingual open-retrieval question answering (QA) systems in 16 typologically diverse languages. In this task, we adapted two large-scale cross-lingual open-retrieval QA datasets in 14 typologically diverse languages, and newly annotated open-retrieval QA data in 2 underrepresented languages: Tagalog and Tamil. Four teams submitted their systems. The best system leveraging iteratively mined diverse negative examples and larger pretrained models achieves 32.2 F1, outperforming our baseline by 4.5 points. The second best system uses entity-aware contextualized representations for document retrieval, and achieves significant improvements in Tamil (20.8 F1), whereas most of the other systems yield nearly zero scores.

cs.CL↗

arXiv2022

DOCmT5: Document-Level Pretraining of Multilingual Language Models

In this paper, we introduce DOCmT5, a multilingual sequence-to-sequence language model pretrained with large scale parallel documents. While previous approaches have focused on leveraging sentence-level parallel data, we try to build a general-purpose pretrained model that can understand and generate long documents. We propose a simple and effective pretraining objective - Document reordering Machine Translation (DrMT), in which the input documents that are shuffled and masked need to be translated. DrMT brings consistent improvements over strong baselines on a variety of document-level generation tasks, including over 12 BLEU points for seen-language-pair document-level MT, over 7 BLEU points for unseen-language-pair document-level MT and over 3 ROUGE-1 points for seen-language-pair cross-lingual summarization. We achieve state-of-the-art (SOTA) on WMT20 De-En and IWSLT15 Zh-En document translation tasks. We also conduct extensive analysis on various factors for document pretraining, including (1) The effects of pretraining data quality and (2) The effects of combining mono-lingual and cross-lingual pretraining. We plan to make our model checkpoints publicly available.

cs.CL↗

arXiv2021

Dialogue State Tracking with a Language Model using Schema-Driven Prompting

Task-oriented conversational systems often use dialogue state tracking to represent the user's intentions, which involves filling in values of pre-defined slots. Many approaches have been proposed, often using task-specific architectures with special-purpose classifiers. Recently, good results have been obtained using more general architectures based on pretrained language models. Here, we introduce a new variation of the language modeling approach that uses schema-driven prompting to provide task-aware history encoding that is used for both categorical and non-categorical slots. We further improve performance by augmenting the prompting with schema descriptions, a naturally occurring source of in-domain knowledge. Our purely generative system achieves state-of-the-art performance on MultiWOZ 2.2 and achieves competitive performance on two other benchmarks: MultiWOZ 2.1 and M2M. The data and code will be available at https://github.com/chiahsuan156/DST-as-Prompting.

cs.CL↗

arXiv2021

KaggleDBQA: Realistic Evaluation of Text-to-SQL Parsers

The goal of database question answering is to enable natural language querying of real-life relational databases in diverse application domains. Recently, large-scale datasets such as Spider and WikiSQL facilitated novel modeling techniques for text-to-SQL parsing, improving zero-shot generalization to unseen databases. In this work, we examine the challenges that still prevent these techniques from practical deployment. First, we present KaggleDBQA, a new cross-domain evaluation dataset of real Web databases, with domain-specific data types, original formatting, and unrestricted questions. Second, we re-examine the choice of evaluation tasks for text-to-SQL parsers as applied in real-life settings. Finally, we augment our in-domain evaluation task with database documentation, a naturally occurring source of implicit domain knowledge. We show that KaggleDBQA presents a challenge to state-of-the-art zero-shot parsers but a more realistic evaluation setting and creative use of associated database documentation boosts their accuracy by over 13.2%, doubling their performance.

cs.CL↗

arXiv2019

Cross-Lingual Transfer Learning for Question Answering

Deep learning based question answering (QA) on English documents has achieved success because there is a large amount of English training examples. However, for most languages, training examples for high-quality QA models are not available. In this paper, we explore the problem of cross-lingual transfer learning for QA, where a source language task with plentiful annotations is utilized to improve the performance of a QA model on a target language task with limited available annotations. We examine two different approaches. A machine translation (MT) based approach translates the source language into the target language, or vice versa. Although the MT-based approach brings improvement, it assumes the availability of a sentence-level translation system. A GAN-based approach incorporates a language discriminator to learn language-universal feature representations, and consequentially transfer knowledge from the source language. The GAN-based approach rivals the performance of the MT-based approach with fewer linguistic resources. Applying both approaches simultaneously yield the best results. We use two English benchmark datasets, SQuAD and NewsQA, as source language data, and show significant improvements over a number of established baselines on a Chinese QA task. We achieve the new state-of-the-art on the Chinese QA dataset.

cs.CL↗

arXiv2019

Mitigating the Impact of Speech Recognition Errors on Spoken Question Answering by Adversarial Domain Adaptation

Spoken question answering (SQA) is challenging due to complex reasoning on top of the spoken documents. The recent studies have also shown the catastrophic impact of automatic speech recognition (ASR) errors on SQA. Therefore, this work proposes to mitigate the ASR errors by aligning the mismatch between ASR hypotheses and their corresponding reference transcriptions. An adversarial model is applied to this domain adaptation task, which forces the model to learn domain-invariant features the QA model can effectively utilize in order to improve the SQA results. The experiments successfully demonstrate the effectiveness of our proposed model, and the results are better than the previous best model by 2% EM score.

cs.CL↗

arXiv2018

ODSQA: Open-domain Spoken Question Answering Dataset

Reading comprehension by machine has been widely studied, but machine comprehension of spoken content is still a less investigated problem. In this paper, we release Open-Domain Spoken Question Answering Dataset (ODSQA) with more than three thousand questions. To the best of our knowledge, this is the largest real SQA dataset. On this dataset, we found that ASR errors have catastrophic impact on SQA. To mitigate the effect of ASR errors, subword units are involved, which brings consistent improvements over all the models. We further found that data augmentation on text-based QA training examples can improve SQA.

cs.CL↗