Selected work

Representative Papers

Evaluating for the long term: Learnings from industry

Aug 08, 2026

Short-term online experiments often fail to accurately predict long-term business outcomes, potentially leading to decisions misaligned with strategic objectives. Drawing on insights from industry expert workshops, this work proposes a design principle for surrogate metrics centered on decision utility rather than solely on unbiasedness, advocating that interpretable, experiment-driven simple surrogates outperform complex black-box models. Through expert consensus synthesis, surrogate metric analysis, and comparative evaluation of experimental versus observational data, the study systematically outlines a methodology for constructing effective surrogates and uncovers a critical relationship between the stability of long-term effects and surrogate validity. While underscoring the irreplaceable value of high-quality long-term experimentation, the research also delineates core challenges and practical guidelines for surrogate learning in real-world settings.

0 citationsRead paper

RARM: Confidence-Gated Progress Reward Modeling for RL in Manipulation

Jun 20, 2026

This work addresses the challenge of reward design in reinforcement learning for robotic manipulation, where sparse rewards provide weak supervision, handcrafted dense rewards suffer from poor generalization, and existing progress-based rewards are often misled by visually plausible but physically infeasible states. To overcome these limitations, the authors propose RARM, a lightweight visual comparator that generates dense, perceptually grounded progress-based rewards from a single successful demonstration. RARM leverages a temporally contrastive model pretrained on unlabeled videos to achieve task- and platform-agnostic reward modeling and incorporates a confidence-gated mechanism to suppress uncertain matches, thereby avoiding spurious positive signals. Evaluated across nine simulated and four real-world manipulation tasks, RARM significantly improves success rates, demonstrating particularly strong performance on long-horizon tasks such as cloth folding.

0 citationsRead paper

SOHET: Sequence Of Heterogeneous Events Transformer with Self-Supervised Pre-Training

Jun 19, 2026

This work addresses the challenges of heterogeneous event stream prediction, where diverse event types and complex temporal dynamics hinder effective modeling. To this end, we propose SOHET, a novel architecture that uniquely integrates event-type-specific tabular encoders with time-type embeddings and leverages either causal or bidirectional Transformers to capture sequential dynamics. We further introduce three self-supervised pretraining objectives tailored for the causal setting, which substantially enhance model generalization and convergence speed. Evaluated on a large-scale fraud detection task from Booking.com, SOHET outperforms existing methods by 5.8%, with pretraining yielding an additional 2.6% performance gain and a 2.4× acceleration in convergence. On the EBES benchmark, SOHET achieves state-of-the-art or competitive results on six out of eight tasks.

0 citationsRead paper

The Significance of Style Diversity in Annotation-Free Synthetic Data Generation

Jun 18, 2026

This work proposes an unsupervised synthetic dialogue generation framework tailored for industrial settings where human-annotated data are scarce, relying solely on intent definitions. To enhance diversity, the approach explicitly incorporates topic and stylistic attributes and introduces two novel post-processing stylization models—Univ and Exam—combined with a large language model–based discriminative filtering mechanism to improve data quality. The study reveals that stylistic diversity has a significantly greater impact on the utility of synthetic data than topic diversity, and that integrating stylistic attributes during generation outperforms post-hoc style transfer. Experimental results demonstrate that the proposed method achieves 93.3% of the performance of models trained on human-annotated data across both industrial and public benchmarks, substantially enhancing the practicality of unlabeled synthetic dialogues.

0 citationsRead paper

On the Limits of Model Merging for Multilinguality in Pre-Training

May 25, 2026

This study investigates whether robust multilingual performance can be achieved by merging monolingual pretrained models. Addressing the limitations of existing approaches in cross-lingual transfer, we systematically compare mixed multilingual pretraining, independent monolingual pretraining, and various model merging strategies. Our experiments reveal that while monolingual models excel in their respective languages, naive merging leads to severe performance collapse, underscoring the critical dependence of successful merging on cross-lingual representational similarity. We further demonstrate, for the first time, that merging mechanisms effective during fine-tuning do not readily generalize to the language-specific pretraining phase. These findings provide crucial theoretical insights and practical warnings for the design and construction of multilingual models.

0 citationsRead paper

Recent publications

Latest Papers

Evaluating for the long term: Learnings from industry

Aug 08, 2026

Short-term online experiments often fail to accurately predict long-term business outcomes, potentially leading to decisions misaligned with strategic objectives. Drawing on insights from industry expert workshops, this work proposes a design principle for surrogate metrics centered on decision utility rather than solely on unbiasedness, advocating that interpretable, experiment-driven simple surrogates outperform complex black-box models. Through expert consensus synthesis, surrogate metric analysis, and comparative evaluation of experimental versus observational data, the study systematically outlines a methodology for constructing effective surrogates and uncovers a critical relationship between the stability of long-term effects and surrogate validity. While underscoring the irreplaceable value of high-quality long-term experimentation, the research also delineates core challenges and practical guidelines for surrogate learning in real-world settings.

0 citationsRead paper

RARM: Confidence-Gated Progress Reward Modeling for RL in Manipulation

Jun 20, 2026

This work addresses the challenge of reward design in reinforcement learning for robotic manipulation, where sparse rewards provide weak supervision, handcrafted dense rewards suffer from poor generalization, and existing progress-based rewards are often misled by visually plausible but physically infeasible states. To overcome these limitations, the authors propose RARM, a lightweight visual comparator that generates dense, perceptually grounded progress-based rewards from a single successful demonstration. RARM leverages a temporally contrastive model pretrained on unlabeled videos to achieve task- and platform-agnostic reward modeling and incorporates a confidence-gated mechanism to suppress uncertain matches, thereby avoiding spurious positive signals. Evaluated across nine simulated and four real-world manipulation tasks, RARM significantly improves success rates, demonstrating particularly strong performance on long-horizon tasks such as cloth folding.

0 citationsRead paper

SOHET: Sequence Of Heterogeneous Events Transformer with Self-Supervised Pre-Training

Jun 19, 2026

This work addresses the challenges of heterogeneous event stream prediction, where diverse event types and complex temporal dynamics hinder effective modeling. To this end, we propose SOHET, a novel architecture that uniquely integrates event-type-specific tabular encoders with time-type embeddings and leverages either causal or bidirectional Transformers to capture sequential dynamics. We further introduce three self-supervised pretraining objectives tailored for the causal setting, which substantially enhance model generalization and convergence speed. Evaluated on a large-scale fraud detection task from Booking.com, SOHET outperforms existing methods by 5.8%, with pretraining yielding an additional 2.6% performance gain and a 2.4× acceleration in convergence. On the EBES benchmark, SOHET achieves state-of-the-art or competitive results on six out of eight tasks.

0 citationsRead paper

The Significance of Style Diversity in Annotation-Free Synthetic Data Generation

Jun 18, 2026

This work proposes an unsupervised synthetic dialogue generation framework tailored for industrial settings where human-annotated data are scarce, relying solely on intent definitions. To enhance diversity, the approach explicitly incorporates topic and stylistic attributes and introduces two novel post-processing stylization models—Univ and Exam—combined with a large language model–based discriminative filtering mechanism to improve data quality. The study reveals that stylistic diversity has a significantly greater impact on the utility of synthetic data than topic diversity, and that integrating stylistic attributes during generation outperforms post-hoc style transfer. Experimental results demonstrate that the proposed method achieves 93.3% of the performance of models trained on human-annotated data across both industrial and public benchmarks, substantially enhancing the practicality of unlabeled synthetic dialogues.

0 citationsRead paper

On the Limits of Model Merging for Multilinguality in Pre-Training

May 25, 2026

This study investigates whether robust multilingual performance can be achieved by merging monolingual pretrained models. Addressing the limitations of existing approaches in cross-lingual transfer, we systematically compare mixed multilingual pretraining, independent monolingual pretraining, and various model merging strategies. Our experiments reveal that while monolingual models excel in their respective languages, naive merging leads to severe performance collapse, underscoring the critical dependence of successful merging on cross-lingual representational similarity. We further demonstrate, for the first time, that merging mechanisms effective during fine-tuning do not readily generalize to the language-specific pretraining phase. These findings provide crucial theoretical insights and practical warnings for the design and construction of multilingual models.

0 citationsRead paper