Selected work

Representative Papers

Learning to Discover at Test Time

Jan 22, 2026

This work proposes TTT-Discover, a novel approach that introduces test-time training (TTT) to scientific discovery by leveraging online reinforcement learning to optimize large language models during inference. Unlike conventional AI methods that rely solely on the generalization of pretrained models and struggle to autonomously identify optimal solutions at test time, TTT-Discover focuses on generating a single high-quality solution rather than improving average performance. Built upon the open-source model OpenAI gpt-oss-120b and augmented with a customized search subroutine and the Tinker API, the method achieves highly efficient and low-cost optimization. It sets new state-of-the-art results across diverse domains—including mathematical theorem proving, GPU kernel design, algorithmic competitions, and single-cell denoising—with all findings validated by domain experts or competition organizers at a cost of only a few hundred dollars per task.

11 citations1 influentialRead paper

Not All LLM Reasoning is Visible in the Chain-of-Thought

Jul 24, 2026

This study investigates whether large language models fully disclose their reasoning processes in outputs—a critical issue for AI interpretability and safety. By introducing semantically irrelevant filler tokens into synthetic reasoning tasks, the authors present the first empirical evidence that state-of-the-art models can exploit such meaningless tokens to perform hidden reasoning, achieving objectives imperceptible to humans from the visible chain of thought. Through controlled experiments manipulating filler tokens and comparing reinforcement learning with supervised fine-tuning across 13 prominent models, the analysis reveals that several models improve accuracy by up to 13 percentage points using fillers. Notably, Claude Opus 4.5 implicitly satisfies modular arithmetic constraints without degrading primary task performance, demonstrating covert computational capabilities absent any explicit trace.

0 citationsRead paper

Addressing the Orchestration Gap in Generalist Robots via Physical Agency

Jul 23, 2026

General-purpose robots struggle to jointly perceive, plan, execute, and recover in complex tasks, as existing end-to-end policies are limited by insufficient reasoning and coordination capabilities. This work proposes Pigey, a physically grounded agent orchestration framework that formally defines and bridges the “orchestration gap” by decoupling high-level task planning from low-level vision-language-action (VLA) policies. Pigey establishes a closed-loop control system that enables goal decomposition, instruction scheduling, outcome verification, and failure recovery. Notably, it achieves substantial performance gains on frozen VLA policies for complex reasoning tasks without requiring additional data or fine-tuning. On the LIBERO-PRO benchmark, the success rate improves from 12.8% to 53.3%, and real-world robotic task success rises from near zero to over 90%.

0 citationsRead paper

ISO: An RLVR-Native Optimization Stack

Jul 21, 2026

This work addresses the limited understanding of how reward feedback efficiently drives weight updates in Verifiable Reward Reinforcement Learning (RLVR). The authors propose a spectral inheritance mechanism that preserves the singular spectrum of model weights while optimizing only the input and output singular frames to learn new behaviors, formalizing this process for the first time as Isospectral Optimization (ISO). This framework enables the fusion of multiple expert models without requiring data, rollouts, gradient updates, or online distillation. By integrating singular value decomposition (SVD) with optimizers such as AdamW and Muon, ISO achieves substantial gains in training efficiency across models ranging from 1.5B to 8B parameters. Notably, Qwen3-8B-Base surpasses the performance of a 270-step baseline in just 100 optimization steps, attaining an accuracy of 0.509.

0 citationsRead paper

Multi-Turn Agentic Scientific Literature Search via Workflow Induction

Jul 01, 2026

Scientific literature retrieval faces significant challenges due to users’ ambiguous, dynamically evolving, and preference-dependent intents. Existing approaches often lack controllability and optimizability owing to their reliance on fixed pipelines or implicit reasoning. To address this, this work proposes PaperPilot, the first framework to introduce an explicit, editable retrieval workflow mechanism. It models retrieval as a directed acyclic graph (DAG) constructed from anchor papers and user queries, supporting operations such as keyword search, citation expansion, filtering, scoring, re-ranking, and evidence extraction. The system iteratively refines both queries and workflow structure through user feedback. Built upon Qwen3.5-9B and integrating supervised imitation learning, preference optimization, and tool calling, PaperPilot achieves substantial improvements in Hit@5 (+19.0→77.0), MRR (+11.9→59.4), and nDCG@10 (+5.7→32.5), while reducing workflow execution errors to 0%.

0 citationsRead paper

Recent publications

Latest Papers

Not All LLM Reasoning is Visible in the Chain-of-Thought

Jul 24, 2026

This study investigates whether large language models fully disclose their reasoning processes in outputs—a critical issue for AI interpretability and safety. By introducing semantically irrelevant filler tokens into synthetic reasoning tasks, the authors present the first empirical evidence that state-of-the-art models can exploit such meaningless tokens to perform hidden reasoning, achieving objectives imperceptible to humans from the visible chain of thought. Through controlled experiments manipulating filler tokens and comparing reinforcement learning with supervised fine-tuning across 13 prominent models, the analysis reveals that several models improve accuracy by up to 13 percentage points using fillers. Notably, Claude Opus 4.5 implicitly satisfies modular arithmetic constraints without degrading primary task performance, demonstrating covert computational capabilities absent any explicit trace.

0 citationsRead paper

Addressing the Orchestration Gap in Generalist Robots via Physical Agency

Jul 23, 2026

General-purpose robots struggle to jointly perceive, plan, execute, and recover in complex tasks, as existing end-to-end policies are limited by insufficient reasoning and coordination capabilities. This work proposes Pigey, a physically grounded agent orchestration framework that formally defines and bridges the “orchestration gap” by decoupling high-level task planning from low-level vision-language-action (VLA) policies. Pigey establishes a closed-loop control system that enables goal decomposition, instruction scheduling, outcome verification, and failure recovery. Notably, it achieves substantial performance gains on frozen VLA policies for complex reasoning tasks without requiring additional data or fine-tuning. On the LIBERO-PRO benchmark, the success rate improves from 12.8% to 53.3%, and real-world robotic task success rises from near zero to over 90%.

0 citationsRead paper

ISO: An RLVR-Native Optimization Stack

Jul 21, 2026

This work addresses the limited understanding of how reward feedback efficiently drives weight updates in Verifiable Reward Reinforcement Learning (RLVR). The authors propose a spectral inheritance mechanism that preserves the singular spectrum of model weights while optimizing only the input and output singular frames to learn new behaviors, formalizing this process for the first time as Isospectral Optimization (ISO). This framework enables the fusion of multiple expert models without requiring data, rollouts, gradient updates, or online distillation. By integrating singular value decomposition (SVD) with optimizers such as AdamW and Muon, ISO achieves substantial gains in training efficiency across models ranging from 1.5B to 8B parameters. Notably, Qwen3-8B-Base surpasses the performance of a 270-step baseline in just 100 optimization steps, attaining an accuracy of 0.509.

0 citationsRead paper

Multi-Turn Agentic Scientific Literature Search via Workflow Induction

Jul 01, 2026

Scientific literature retrieval faces significant challenges due to users’ ambiguous, dynamically evolving, and preference-dependent intents. Existing approaches often lack controllability and optimizability owing to their reliance on fixed pipelines or implicit reasoning. To address this, this work proposes PaperPilot, the first framework to introduce an explicit, editable retrieval workflow mechanism. It models retrieval as a directed acyclic graph (DAG) constructed from anchor papers and user queries, supporting operations such as keyword search, citation expansion, filtering, scoring, re-ranking, and evidence extraction. The system iteratively refines both queries and workflow structure through user feedback. Built upon Qwen3.5-9B and integrating supervised imitation learning, preference optimization, and tool calling, PaperPilot achieves substantial improvements in Hit@5 (+19.0→77.0), MRR (+11.9→59.4), and nDCG@10 (+5.7→32.5), while reducing workflow execution errors to 0%.

0 citationsRead paper

Sketched Linear Contrastive Learning: Approximation, Optimization, and Statistical Scaling

Jun 25, 2026

This work investigates the scaling laws of contrastive learning under varying model sizes, data volumes, and computational budgets, addressing a critical theoretical gap in the field. Focusing on linear contrastive learning with paired Gaussian latent variables and observed two-view sketches, the study employs full-batch gradient descent to train a bilinear scoring function. By introducing a Gaussian-negative quadratic surrogate loss, a power-law spectral alignment assumption, and a contrastive source condition, it establishes—for the first time—an explicit trivariate statistical scaling law that jointly accounts for sketch dimensionality, sample size, and optimization duration. This law reveals the distinct influence of two-view interactions on optimization dynamics and noise sensitivity. Through a bias–variance–cross-term decomposition, the authors derive a risk expression comprising irreducible risk, approximation error, optimization bias, and variance, offering a principled theoretical foundation for balancing model capacity, data quantity, and computational effort.

0 citationsRead paper