Institution profile

Shanghai Qi Zhi Institute

Academic institutionasia · cn
Official website
Research library170linked papers
Opportunities0open roles
Selected work

Representative Papers

Cumulative Reasoning with Large Language Models

Aug 08, 2023arXiv.org

Large language models (LLMs) remain constrained in complex logical and mathematical reasoning. This paper introduces Cumulative Reasoning (CR), a novel framework that emulates human incremental thinking by decomposing problems into subtasks and iteratively accumulating and reusing intermediate propositions and execution results—without external retrieval or internet access. CR integrates symbolic reasoning chain construction, iterative prompt engineering, and an internal code execution environment to enable pure model-intrinsic reasoning leaps. Experiments demonstrate substantial gains: 98.04% accuracy on FOLIO (+9.3%), 98% on Game of 24 (+24%), +4.2% overall on MATH and up to +43% on its hardest subset. A code-only CR variant achieves 72.2%, significantly outperforming PAL and PoT (+38.8%). This work is the first to systematically propose and empirically validate an intrinsic, cumulative, iterative reasoning paradigm.

59 citations5 influentialRead paper

On the Diagram of Thought

Sep 16, 2024arXiv.org

This work addresses the key limitation in large language models (LLMs) where iterative reasoning is constrained to linear chains or trees, hindering simultaneous support for diverse reasoning paths and logical consistency. To this end, we propose the Diagram of Thought (DoT) framework, which unifies iterative reasoning within a single model via a directed acyclic graph (DAG)—where nodes represent propositions alongside their critique, refinement, and verification—and enables natural-language-feedback-driven self-optimization. Methodologically, DoT introduces a novel formal foundation grounded in Topos theory and a role-specific token-augmented autoregressive prediction mechanism, allowing seamless alternation between generative and critical roles within one model. Experiments demonstrate that DoT significantly improves both reasoning consistency and path exploration capability—without requiring multi-model collaboration or external controllers—establishing a new paradigm for scalable, principle-driven design of reasoning-specialized models.

9 citations1 influentialRead paper

Data Scaling Laws in Imitation Learning for Robotic Manipulation

Oct 24, 2024arXiv.org

This study investigates data scaling laws in robotic imitation learning, addressing whether a single policy can achieve zero-shot generalization to unseen environments and objects of the same category using limited yet high-quality demonstration data. We propose a “diversity-first, demonstration-count-threshold-driven” data collection paradigm for efficient real-world robot data acquisition. We empirically uncover, for the first time, a power-law relationship between the number of training environments/objects and zero-shot performance. Evaluation is conducted under a rigorous, realistic cross-environment and cross-object benchmark using behavior cloning. With only half-a-day of real-robot demonstrations collected by four human operators—totaling ~4 hours—we achieve approximately 90% success rates on two distinct manipulation tasks across unseen environments and unseen objects. This demonstrates substantial improvements in data efficiency and generalization robustness, establishing scalable, practical foundations for deploying imitation learning in diverse real-world settings.

7 citationsRead paper

Hiking in the Wild: A Scalable Perceptive Parkour Framework for Humanoids

Jan 12, 2026

This work addresses the challenge of achieving robust, high-speed locomotion for humanoid robots in complex, unstructured environments, where reliance on proprioception alone is insufficient and incorporating exteroception often introduces state estimation drift and poor training scalability. To overcome these limitations, the authors propose an end-to-end perception-to-action framework that directly maps raw depth images and proprioceptive inputs to joint actions without explicit state estimation. The approach integrates terrain edge detection, a foot-volume-based foothold safety mechanism, and a flat-region sampling strategy to enhance training stability and deployment safety. Implemented within a single-stage reinforcement learning architecture, the method enables a full-scale humanoid robot to traverse challenging terrains at speeds up to 2.5 m/s. The code is open-sourced and supports real-world deployment with minimal hardware modifications.

4 citationsRead paper

Meta Prompting for AI Systems

Nov 20, 2023

Large language models (LLMs) suffer from low prompting efficiency and poor generalization in complex reasoning and dynamic data interaction tasks. Method: We propose Meta Prompting—a novel prompting paradigm that systematically integrates type theory and category theory into prompt engineering, enabling compositional, formally verifiable prompt structures. It supports task decomposition, recursive self-prompting, and metaprogramming-style prompt evolution—without any parameter fine-tuning. Contribution/Results: Our framework establishes the first formal theoretical foundation for prompting, significantly enhancing zero-shot high-order reasoning and open-ended interactive capabilities. Experiments show Qwen-72B achieves 46.3% accuracy on MATH and 83.5% on GSM8K—surpassing same-scale supervised fine-tuned models—while GPT-4 attains 100% success on the Game of 24. These results demonstrate substantial improvements in zero-shot performance across rigorous mathematical and procedural reasoning benchmarks.

4 citationsRead paper
Recent publications

Latest Papers