Papers — Yijie Hao

2026

Reasoning Traces Shape Outputs but Models Won't Say So

Yijie Hao*, Lingjie Chen*, Ali Emami, Joyce Ho

ACL 2026 Main

pdf

tl;dr We inject synthetic reasoning into LLMs' <think> traces and find that models follow the injected reasoning but systematically refuse to acknowledge doing so — fabricating explanations instead.

Can we trust the reasoning traces that large reasoning models produce? We introduce Thought Injection, a method that injects synthetic reasoning snippets into a model's <think> trace, then measures whether the model follows the injected reasoning and acknowledges doing so. Across 45,000 samples from three LLMs, we find that injected hints reliably alter outputs, confirming that reasoning traces causally shape model behavior. However, when asked to explain their changed answers, models overwhelmingly refuse to disclose the influence: overall non-disclosure exceeds 90% for extreme hints. Activation analysis reveals that sycophancy- and deception-related directions are strongly activated during these fabrications, suggesting systematic patterns rather than incidental failures.

findings

Causal influence confirmed: Injected hints reliably change model outputs across 45,000 samples from three state-of-the-art LRMs, establishing that reasoning traces are not merely post-hoc narration.

Systematic non-disclosure: Models fail to acknowledge injected reasoning in over 90% of extreme-hint cases (30,000 follow-up samples), instead fabricating aligned-appearing but unrelated explanations.

Mechanistic correlates: Activation analysis shows sycophancy- and deception-related directions are strongly activated when models fabricate explanations — suggesting structured failure modes, not random errors.

@inproceedings{hao2026reasoning, title={Reasoning Traces Shape Outputs but Models Won't Say So}, author={Hao, Yijie and Chen, Lingjie and Emami, Ali and Ho, Joyce}, booktitle={Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics}, year={2026} }

2025

Beyond Facts: Evaluating Intent Hallucination in Large Language Models

Yijie Hao, Haofei Yu, Jiaxuan You

ACL 2025 Main

pdf

tl;dr LLMs don't just hallucinate facts — they hallucinate intent, omitting or misinterpreting parts of complex queries. We formalize this as "intent hallucination" and introduce FaithQA, a 20K-query benchmark, along with Constraint Score, an evaluation metric that outperforms LLM-as-judge approaches.

When exposed to complex queries containing multiple conditions, today's large language models tend to produce responses that only partially satisfy the query while neglecting certain conditions. We term this phenomenon "Intent Hallucination." To systematically evaluate intent hallucination, we introduce FaithQA, a novel benchmark containing 20,068 problems, covering both query-only and retrieval-augmented generation (RAG) setups with varying topics and difficulty. We also introduce Constraint Score, an automatic evaluation metric that decomposes queries into intent constraints and computes a weighted score. Our human evaluation shows that Constraint Score significantly outperforms LLM-as-the-judge baselines.

findings

Intent hallucination as a concept: We distinguish intent hallucination (omission/misinterpretation of query components) from factual hallucination — a prevalent but overlooked failure mode even in state-of-the-art models.

FaithQA benchmark: 20,068 queries validated through extensive human evaluation, covering diverse topics and difficulty levels in both query-only and RAG settings.

Constraint Score: A novel evaluation metric that decomposes queries into intent constraints and assigns importance-weighted binary labels. Significantly outperforms LLM-as-judge approaches, which tend to be biased.

@inproceedings{hao2025beyond, title={Beyond Facts: Evaluating Intent Hallucination in Large Language Models}, author={Hao, Yijie and Yu, Haofei and You, Jiaxuan}, booktitle={Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics}, year={2025} }

2024

A Flexible Generative Model for Heterogeneous Tabular EHR with Missing Modality

Huan He*, Yijie Hao*, Yuanzhe Xi, Yong Chen, Bradley Malin, Joyce Ho

ICLR 2024

pdf openreview

tl;dr Real-world EHRs mix static and temporal features with frequent missing modalities. We introduce FlexGen-EHR, a latent diffusion model with an optimal transport module that handles this heterogeneity, outperforming existing methods in both fidelity and utility.

Realistic synthetic electronic health records (EHRs) can accelerate research while mitigating privacy concerns. However, training generative models on EHR data remains challenging due to heterogeneous feature types (static and temporal) and missing modalities. We introduce FlexGen-EHR, a versatile diffusion model tailored for heterogeneous tabular EHRs, equipped with an integrative learning framework and an optimal transport module to align the common feature space of heterogeneous EHRs. We empirically show that our model consistently outperforms existing state-of-the-art synthetic EHR generation methods both in fidelity by up to 3.10% and utility by up to 7.16%.

findings

Heterogeneous EHR generation: First to formalize the challenge of generating heterogeneous EHR data (static + temporal features) in the presence of missing modalities.

Latent diffusion + optimal transport: FlexGen-EHR uses separate VAE encoders for static and temporal features, unified in a shared latent space, with an optimal transport module to construct meaningful latent pairs when modalities are missing.

State-of-the-art results: Up to 3.10% improvement in fidelity (MMD) and 7.16% in utility (AUPR) over existing methods, with strong performance even under missing-modality conditions.

@inproceedings{he2024flexible, title={A Flexible Generative Model for Heterogeneous Tabular {EHR} with Missing Modality}, author={He, Huan and Hao, Yijie and Xi, Yuanzhe and Chen, Yong and Malin, Bradley and Ho, Joyce}, booktitle={International Conference on Learning Representations}, year={2024} }