arXiv subjects

Yang Sheng

Publications and source records attributed to Yang Sheng.

7 recordsLinked to original sources

arXiv2026

Circuit Claims Depend on What Is Extracted and How It Is Compared

Circuit extraction identifies a small set of model components whose presence preserves a target behavior under ablation, and the resulting circuit is often read as the mechanism behind that behavior. We argue that this reading is under-determined: preserving behavior does not single out one circuit, because the claim it supports depends on which circuit is reported and how two circuits are compared. We make this concrete in a synthetic Lean tactic-prediction benchmark -- predicting the next step of a proof -- where fixed proof rules with randomized surface form let differences between extracted circuits be attributed to these choices rather than to the task. Across dense and weight-sparse checkpoints (most weights constrained to zero) of the same transformer, evaluated on atomic (single-rule) and compositional (multi-rule) proofs, we vary which extracted object is reported (a compact prediction-preserving circuit, a broader graph that also keeps surrounding read, write, and routing structure, or the smallest subgraph meeting a post-ablation loss threshold), and whether each attention head's query and key are represented jointly or separately. Exact component-to-component edge overlap is low and sensitive to these choices, at times dropping to a random baseline, while two coarser summaries stay stable: the set of selected attention heads, and the circuit-size ranking of conditions that differ in which supervised checkpoint initializes reinforcement learning (RL). The largest accuracy gains from RL on compositional proofs come with the most structure beyond the atomic circuits. A circuit-level claim is therefore well defined only once one states which circuit is reported, the pruning threshold used to extract it, and the level at which circuits are compared. We distill these requirements into a reporting practice for circuit-extraction studies.

cs.LG↗

arXiv2026

FormalJudge: A Neuro-Symbolic Paradigm for Agentic Oversight

As LLM-based agents increasingly operate in high-stakes domains with real-world consequences, ensuring their behavioral safety becomes paramount. The dominant oversight paradigm, LLM-as-a-Judge, faces a fundamental dilemma: how can probabilistic systems reliably supervise other probabilistic systems without inheriting their failure modes? We argue that formal verification offers a principled escape from this dilemma, yet its adoption has been hindered by a critical bottleneck: the translation from natural language requirements to formal specifications. This paper bridges this gap by proposing , a neuro-symbolic framework that employs a bidirectional Formal-of-Thought architecture: LLMs serve as specification compilers that top-down decompose high-level human intent into atomic, verifiable constraints, then bottom-up prove compliance using Dafny specifications and Z3 Satisfiability modulo theories solving, which produces mathematical guarantees rather than probabilistic scores. We validate across three benchmarks spanning behavioral safety, multi-domain constraint adherence, and agentic upward deception detection. Experiments on 7 agent models demonstrate that achieves an average improvement of 16.6% over LLM-as-a-Judge baselines, enables weak-to-strong generalization where a 7B judge achieves over 90% accuracy detecting deception from 72B agents, and provides near-linear safety improvement through iterative refinement.

cs.AI↗

arXiv2025

Zero-Shot Large Language Model Agents for Fully Automated Radiotherapy Treatment Planning

Radiation therapy treatment planning is an iterative, expertise-dependent process, and the growing burden of cancer cases has made reliance on manual planning increasingly unsustainable, underscoring the need for automation. In this study, we propose a workflow that leverages a large language model (LLM)-based agent to navigate inverse treatment planning for intensity-modulated radiation therapy (IMRT). The LLM agent was implemented to directly interact with a clinical treatment planning system (TPS) to iteratively extract intermediate plan states and propose new constraint values to guide inverse optimization. The agent's decision-making process is informed by current observations and previous optimization attempts and evaluations, allowing for dynamic strategy refinement. The planning process was performed in a zero-shot inference setting, where the LLM operated without prior exposure to manually generated treatment plans and was utilized without any fine-tuning or task-specific training. The LLM-generated plans were evaluated on twenty head-and-neck cancer cases against clinical manual plans, with key dosimetric endpoints analyzed and reported. The LLM-generated plans achieved comparable organ-at-risk (OAR) sparing relative to clinical plans while demonstrating improved hot spot control (Dmax: 106.5% vs. 108.8%) and superior conformity (conformity index: 1.18 vs. 1.39 for boost PTV; 1.82 vs. 1.88 for primary PTV). This study demonstrates the feasibility of a zero-shot, LLM-driven workflow for automated IMRT treatment planning in a commercial TPS. The proposed approach provides a generalizable and clinically applicable solution that could reduce planning variability and support broader adoption of AI-based planning strategies.

physics.med-ph↗

arXiv2024

Large Language Models Can Better Understand Knowledge Graphs Than We Thought

When we integrate factual knowledge from knowledge graphs (KGs) into large language models (LLMs) to enhance their performance, the cost of injection through training increases with the scale of the models. Consequently, there is significant interest in developing prompt strategies that effectively incorporate KG information into LLMs. However, the community has not yet comprehensively understood how LLMs process and interpret KG information in different input formats and organizations within prompts, and researchers often rely on trial and error. To address this gap, we design extensive experiments to empirically study LLMs' comprehension of different KG prompts. At the literal level, we reveal LLMs' preferences for various input formats (from linearized triples to fluent natural language text). At the attention distribution level, we discuss the underlying mechanisms driving these preferences. We then investigate how the organization of structured knowledge impacts LLMs and evaluate LLMs' robustness in processing and utilizing KG information in practical scenarios. Our experiments show that (1) linearized triples are more effective than fluent NL text in helping LLMs understand KG information and answer fact-intensive questions; (2) Different LLMs exhibit varying preferences for different organizational formats of triples; (3) LLMs with larger scales are more susceptible to noisy, incomplete subgraphs.

cs.CL↗

arXiv2023

A Radiomics-Incorporated Deep Ensemble Learning Model for Multi-Parametric MRI-based Glioma Segmentation

We developed a deep ensemble learning model with a radiomics spatial encoding execution for improved glioma segmentation accuracy using multi-parametric MRI (mp-MRI). This model was developed using 369 glioma patients with a 4-modality mp-MRI protocol: T1, contrast-enhanced T1 (T1-Ce), T2, and FLAIR. In each modality volume, a 3D sliding kernel was implemented across the brain to capture image heterogeneity: fifty-six radiomic features were extracted within the kernel, resulting in a 4th order tensor. Each radiomic feature can then be encoded as a 3D image volume, namely a radiomic feature map (RFM). PCA was employed for data dimension reduction and the first 4 PCs were selected. Four deep neural networks as sub-models following the U-Net architecture were trained for the segmenting of a region-of-interest (ROI): each sub-model utilizes the mp-MRI and 1 of the 4 PCs as a 5-channel input for a 2D execution. The 4 softmax probability results given by the U-net ensemble were superimposed and binarized by Otsu method as the segmentation result. Three ensemble models were trained to segment enhancing tumor (ET), tumor core (TC), and whole tumor (WT). The adopted radiomics spatial encoding execution enriches the image heterogeneity information that leads to the successful demonstration of the proposed deep ensemble model, which offers a new tool for mp-MRI based medical image segmentation.

q-bio.QM↗

arXiv2020

An Artificial Intelligence-Driven Agent for Real-Time Head-and-Neck IMRT Plan Generation using Conditional Generative Adversarial Network (cGAN)

Purpose: To develop an Artificial Intelligence (AI) agent for fully-automated rapid head and neck (H&N) IMRT plan generation without time-consuming inverse planning.$$$$ Methods: This AI agent was trained using a conditional Generative Adversarial Network architecture. The generator, PyraNet, is a novel Deep Learning network that implements 28 classic ResNet blocks in pyramid-like concatenations. The discriminator is a customized 4-layer DenseNet. The AI agent first generates customized 2D projections at 9 template beam angles from 3D CT volume and structures of a patient. These projections are then stacked as 4D inputs of PyraNet, from which 9 radiation fluence maps are generated simultaneously. Finally, the predicted fluence maps are imported into a commercial treatment planning system (TPS) for plan integrity checks. The AI agent was built and tested upon 231 oropharyngeal plans from a TPS plan library. Only the primary plans in the sequential boost regime were studied. A customized Harr wavelet loss was adopted for fluence map comparison. Isodose distributions in test AI plans and TPS plans were qualitatively evaluated. Key dosimetric metrics were statistically compared.$$$$ Results: All test AI plans were successfully generated. Isodose gradients outside of PTV in AI plans were comparable with TPS plans. After PTV coverage normalization, $D_{mean}$ of parotids and oral cavity in AI plans and TPS plans were comparable without statistical significance. AI plans achieved comparable $D_{max}$ at 0.01cc of brainstem and cord+5mm without clinically relevant differences, but body $D_{max}$ was higher than the TPS plan results. The AI agent needs ~3s per case to predict fluence maps.$$$$ Conclusions: The developed AI agent can generate H&N IMRT plans with satisfying dosimetry quality. With rapid and fully automated implementation, it holds great potential for clinical applications.

physics.med-ph↗

arXiv2020

An interpretable planning bot for pancreas stereotactic body radiation therapy

Pancreas stereotactic body radiotherapy treatment planning requires planners to make sequential, time consuming interactions with the treatment planning system (TPS) to reach the optimal dose distribution. We seek to develop a reinforcement learning (RL)-based planning bot to systematically address complex tradeoffs and achieve high plan quality consistently and efficiently. The focus of pancreas SBRT planning is finding a balance between organs-at-risk sparing and planning target volume (PTV) coverage. Planners evaluate dose distributions and make planning adjustments to optimize PTV coverage while adhering to OAR dose constraints. We have formulated such interactions between the planner and the TPS into a finite-horizon RL model. First, planning status features are evaluated based on human planner experience and defined as planning states. Second, planning actions are defined to represent steps that planners would commonly implement to address different planning needs. Finally, we have derived a reward system based on an objective function guided by physician-assigned constraints. The planning bot trained itself with 48 plans augmented from 16 previously treated patients and generated plans for 24 cases in a separate validation set. All 24 bot-generated plans achieve similar PTV coverages compared to clinical plans while satisfying all clinical planning constraints. Moreover, the knowledge learned by the bot can be visualized and interpreted as consistent with human planning knowledge, and the knowledge maps learned in separate training sessions are consistent, indicating reproducibility of the learning process.

physics.med-ph↗