Thinh Pham

Quyet V. Do, Thinh Pham, Nguyen Nguyen et al.

We study a pipeline that curates reasoning data from initial structured data for improving long-context reasoning in large language models (LLMs). Our approach, $π^2$, constructs high-quality reasoning data through rigorous QA curation: 1) extracting and expanding tables from Wikipedia, 2) from the collected tables and relevant context, generating realistic and multi-hop analytical reasoning questions whose answers are automatically determined and verified through dual-path code execution, and 3) back-translating step-by-step structured reasoning traces as solutions of QA pairs given realistic web-search context. Supervised fine-tuning with \textsc{\small{gpt-oss-20b}} and \textsc{\small{Qwen3-4B-Instruct-2507}} on $π^2$ yields consistent improvements across four long-context reasoning benchmarks and our alike $π^2$-Bench, with average absolute accuracy gains of +4.3% and +2.7% respectively. Notably, our dataset facilitates self-distillation, where \textsc{\small{gpt-oss-20b}} even improves its average performance by +4.4% with its own reasoning traces, demonstrating $π^2$'s usefulness. Our code, data, and models are open-source at https://github.com/vt-pi-squared/pi-squared.

Han Wang, David Wan, Hyunji Lee et al.

Motivated by the underspecified, multi-hop nature of search queries and the multimodal, heterogeneous, and often conflicting nature of real-world web results, we introduce MERRIN (Multimodal Evidence Retrieval and Reasoning in Noisy Web Environments), a human-annotated benchmark for evaluating search-augmented agents. MERRIN measures AI agents' ability to identify relevant modalities, retrieve multimodal evidence, and perform multi-hop reasoning over noisy web sources. It differs from prior work in three important aspects: (1) using natural language queries without explicit modality cues, (2) incorporating underexplored modalities such as video and audio, and (3) requiring the retrieval of complex, often noisy or conflicting multimodal evidence during web search. We evaluate diverse search agents powered by ten models, including strong closed-source models (e.g., GPT-5.4-mini, Gemini 3/3.1 Flash/Pro) and open-weight models (Qwen3-4B/30B/235B), across three search settings (no search, native search, and agentic search). Our results show that MERRIN is highly challenging: the average accuracy across all agents is 22.3%, with the best-performing agent reaching only 40.1%. We further observe that while stronger agents like Gemini Deep Research achieve higher performance, gains are modest due to over-exploration; they take more steps and use more tools, but are often distracted by conflicting or partially relevant web content, leading to incorrect answers. Compared to humans, these agents consume more resources yet achieve lower accuracy, largely due to inefficient source selection and an overreliance on text modalities. These findings highlight the need for search agents capable of robust search and reasoning across diverse modalities in noisy web environments, making MERRIN a valuable testbed for evaluating such capabilities.

Thinh Pham, Nguyen Nguyen, Pratibha Zunjare et al.

We introduce SealQA, a new challenge benchmark for evaluating SEarch-Augmented Language models on fact-seeking questions where web search yields conflicting, noisy, or unhelpful results. SealQA comes in three flavors: (1) Seal-0 (main) and (2) Seal-Hard, which assess factual accuracy and reasoning capabilities, with Seal-0 focusing on the most challenging questions where chat models (e.g., GPT-4.1) typically achieve near-zero accuracy; and (3) LongSeal, which extends SealQA to test long-context, multi-document reasoning in "needle-in-a-haystack" settings. Our evaluation reveals critical limitations in current models: Even frontier LLMs perform poorly across all SealQA flavors. On Seal-0, frontier agentic models equipped with tools like o3 and o4-mini achieve only 17.1% and 6.3% accuracy, respectively, at their best reasoning efforts. We find that advanced reasoning models such as DeepSeek-R1-671B and o3-mini are highly vulnerable to noisy search results. Notably, increasing test-time compute does not yield reliable gains across o3-mini, o4-mini, and o3, with performance often plateauing or even declining early. Additionally, while recent models are less affected by the "lost-in-the-middle" issue, they still fail to reliably identify relevant documents in LongSeal when faced with numerous distractors. To facilitate future work, we release SealQA at huggingface.co/datasets/vtllms/sealqa.

Linh The Nguyen, Thinh Pham, Dat Quoc Nguyen

We present XPhoneBERT, the first multilingual model pre-trained to learn phoneme representations for the downstream text-to-speech (TTS) task. Our XPhoneBERT has the same model architecture as BERT-base, trained using the RoBERTa pre-training approach on 330M phoneme-level sentences from nearly 100 languages and locales. Experimental results show that employing XPhoneBERT as an input phoneme encoder significantly boosts the performance of a strong neural TTS model in terms of naturalness and prosody and also helps produce fairly high-quality speech with limited training data. We publicly release our pre-trained XPhoneBERT with the hope that it would facilitate future research and downstream TTS applications for multiple languages. Our XPhoneBERT model is available at https://github.com/VinAIResearch/XPhoneBERT

Morgan Lee, Artem Frenk, Eamon Worden et al.

Knowledge Tracing (KT) has been an established problem in the educational data mining field for decades, and it is commonly assumed that the underlying learning process being modeled remains static. Given the ever-changing landscape of online learning platforms (OLPs), we investigate how concept drift and changing student populations can impact student behavior within an OLP through testing model performance both within a single academic year and across multiple academic years. Four well-studied KT models were applied to five academic years of data to assess how susceptible KT models are to concept drift. Through our analysis, we find that all four families of KT models can exhibit degraded performance, Bayesian Knowledge Tracing (BKT) remains the most stable KT model when applied to newer data, while more complex, attention based models lose predictive power significantly faster.

Thinh Pham, Dat Quoc Nguyen

Profile-based intent detection and slot filling are important tasks aimed at reducing the ambiguity in user utterances by leveraging user-specific supporting profile information. However, research in these two tasks has not been extensively explored. To fill this gap, we propose a joint model, namely JPIS, designed to enhance profile-based intent detection and slot filling. JPIS incorporates the supporting profile information into its encoder and introduces a slot-to-intent attention mechanism to transfer slot information representations to intent detection. Experimental results show that our JPIS substantially outperforms previous profile-based models, establishing a new state-of-the-art performance in overall accuracy on the Chinese benchmark dataset ProSLU.

Khang Tran, Hieu Cao, Thinh Pham et al.

Regression is essential across many domains but remains challenging in high-dimensional settings, where existing methods often lose spatial structure or demand heavy storage. In this work, we address the problem of matrix-valued regression, where each sample is naturally represented as a matrix. We propose MELCOT, a hybrid model that integrates a classical machine learning-based Marginal Estimation (ME) block with a deep learning-based Learnable-Cost Optimal Transport (LCOT) block. The ME block estimates data marginals to preserve spatial information, while the LCOT block learns complex global features. This design enables MELCOT to inherit the strengths of both classical and deep learning methods. Extensive experiments across diverse datasets and domains demonstrate that MELCOT consistently outperforms all baselines while remaining highly efficient.

Thinh Pham, Chi Tran, Dat Quoc Nguyen

The research study of detecting multiple intents and filling slots is becoming more popular because of its relevance to complicated real-world situations. Recent advanced approaches, which are joint models based on graphs, might still face two potential issues: (i) the uncertainty introduced by constructing graphs based on preliminary intents and slots, which may transfer intent-slot correlation information to incorrect label node destinations, and (ii) direct incorporation of multiple intent labels for each token w.r.t. token-level intent voting might potentially lead to incorrect slot predictions, thereby hurting the overall performance. To address these two issues, we propose a joint model named MISCA. Our MISCA introduces an intent-slot co-attention mechanism and an underlying layer of label attention mechanism. These mechanisms enable MISCA to effectively capture correlations between intents and slot labels, eliminating the need for graph construction. They also facilitate the transfer of correlation information in both directions: from intents to slots and from slots to intents, through multiple levels of label-specific representations, without relying on token-level intent information. Experimental results show that MISCA outperforms previous models, achieving new state-of-the-art overall accuracy performances on two benchmark datasets MixATIS and MixSNIPS. This highlights the effectiveness of our attention mechanisms.