Papers by Xiangyu Peng

15 papers
Guiding Neural Story Generation with Reader Models (2022.findings-emnlp)

Copied to clipboard

Challenge: Existing systems that generate narratives with neural language models require substantial knowledge engineering of logical constraints, limiting their generality.
Approach: They propose a framework in which a reader model is used to reason about the storyshould progress.
Outcome: The proposed model outperforms baseline models in plot plausibility and staying on topic.
PaCoRe: Learning to Scale Test-Time Compute with Parallel Coordinated Reasoning (2026.acl-long)

Copied to clipboard

Challenge: Parallel Coordinated Reasoning (PaCoRe) overcomes a central limitation of contemporary language models: their inability to scale test-time compute (TTC) far beyond sequential reasoning under a fixed context window.
Approach: They propose a training-and-inference framework to overcome a central limitation of language models: their inability to scale test-time compute (TTC) under a fixed context window.
Outcome: The proposed model scales to multi-million-token effective TTC without exceeding context limits.
Don’t Stop Early: Scalable Enterprise Deep Research with Controlled Information Flow and Evidence-Aware Termination (2026.acl-industry)

Copied to clipboard

Challenge: Enterprise deep research systems fail to produce decision-ready reports due to uneven information coverage, context explosion, and premature stopping.
Approach: They propose a scalable Enterprise Deep Research (EDR) architecture that decomposes requests into coverage-driven objectives via outline generation with reflection and localizes context with dependency-guided execution and explicit information sharing.
Outcome: The proposed system achieves the strongest overall performance compared with competitive deep-research baselines on internal sales enablement tasks and the public DeepResearch Bench benchmark.
HiddenDetect: Detecting Jailbreak Attacks against Multimodal Large Language Models via Monitoring Hidden States (2025.acl-long)

Copied to clipboard

Challenge: Existing studies focus on posthoc alignment techniques, but the underlying safety mechanisms within LVLMs remain unexplored.
Approach: They propose a tuning-free framework that leverages internal activations to enhance safety.
Outcome: The proposed framework outperforms state-of-the-art methods in detecting jailbreak attacks against large vision-language models.
Inferring the Reader: Guiding Automated Story Generation with Commonsense Reasoning (2022.findings-emnlp)

Copied to clipboard

Challenge: Existing methods to automate story generation focus on single-character stories and lack basiccommonsense reasoning.
Approach: They propose a commonsense-inference Augmentedneural StoryTelling framework that introduces commonsensical reasoning into the story generation process.
Outcome: The proposed method produces significantly more coherent, on-topic, enjoyable andfluent stories than existing models in both the single-character and two-character settings.
Logic-Consistency Text Generation from Semantic Parses (2021.findings-acl)

Copied to clipboard

Challenge: Text generation from semantic parses is challenging due to the complexity of the inner logic and the lack of automatic evaluation metrics for logic consistency.
Approach: They propose a framework for logic consistent text generation from semantic parses that employs iterative training procedures and quality control.
Outcome: The proposed framework enhances logic consistency and human evaluation on two benchmark datasets.
Benchmarking Deep Search over Heterogeneous Enterprise Data (2025.emnlp-industry)

Copied to clipboard

Challenge: Existing methods struggle to conduct deep searches and retrieve all necessary evidence.
Approach: They propose a benchmark for evaluating deep search, a retrieval-augmented generation that requires source-aware, multi-hop reasoning over diverse, sparsed, but related sources.
Outcome: The proposed benchmarks show that even the best-performing agentic RAG methods achieve an average performance score of 32.96 on the benchmark.
Turning Conversations into Workflows: A Framework to Extract and Evaluate Dialog Workflows for Service AI Agents (2025.findings-acl)

Copied to clipboard

Challenge: Existing workflow extraction methods for service agents are time-consuming and outdated, causing inconsistent and inconsistent results.
Approach: They propose a framework for extracting and evaluating dialog workflows from historical interactions.
Outcome: The proposed framework improves workflow extraction by 12.16% over baseline.
Unanswerability Evaluation for Retrieval Augmented Generation (2025.acl-long)

Copied to clipboard

Challenge: Existing evaluation frameworks for retrieval-augmented generation (RAG) systems focus on answerable queries, but ignore the importance of appropriately rejecting unanswerable requests.
Approach: They propose a framework to evaluate whether retrieval-augmented generation systems handle unanswerable queries specific to a given knowledge base.
Outcome: The proposed framework synthesizes diverse and challenging queries for any given knowledge base and evaluates them with unanswered ratio and acceptable ratio metrics.
Exploring Reasoning Reward Model for Agents (2026.findings-acl)

Copied to clipboard

Challenge: Existing methods for agentic reinforcement learning rely on sparse outcome-based reward for training, leading to suboptimal results.
Approach: They propose an agent-based reward model that produces structured feedback for agentic trajectories, including an explicit reasoning trace and a focused critique.
Outcome: The proposed model produces structured feedback for agentic trajectories including an explicit reasoning trace, a focused critique, and an overall score that evaluates process performance.
BalanceSFT: Improving LLM Function Calling with Balanced Training Signals and Data Hardness (2026.findings-acl)

Copied to clipboard

Challenge: Currently, Supervised Fine-Tuning (SFT) is the prevailing method for equipping Large Language Models (LLMs) with function calling capabilities, but its effectiveness is often compromised by two challenges: 1) lengthy Chain-of-Thought (CoT) reasoning tokens dominate training signals over concise function calls in the learning objective; 2) scarcity of hard training examples.
Approach: They propose a framework that uses a self-adjusted signal balancing loss and a hard data re-sampling strategy to selectively generate new, high-quality complex data guided by model errors.
Outcome: The proposed framework surpasses state-of-the-art models like GPT-5 in function calling performance.
T2R-BENCH: A Benchmark for Real World Table-to-Report Task (2025.emnlp-main)

Copied to clipboard

Challenge: Existing table benchmarks lack the capacity to adequately assess the practical application of table reasoning in industrial applications.
Approach: They propose a bilingual table-to-report task and a table-based benchmark to assess the quality of table reasoning.
Outcome: The proposed task is based on a bilingual benchmark with 457 industrial tables and evaluation criteria to measure the quality of report generation.
Automatic Bug Detection in LLM-Powered Text-Based Games Using LLMs (2024.findings-acl)

Copied to clipboard

Challenge: Advancements in large language models (LLMs) are revolutionizing interactive game design, but they may exhibit flaws such as hallucinations, forgetfulness, or misinterpretation of prompts.
Approach: They propose a method for automatically identifying LLM bugs from player game logs . their method surpasses unstructured bug-catching methods and fills the gap .
Outcome: The proposed method surpasses unstructured bug-catching methods and fills the gap in detection of logical and design flaws.
Stepwise Reasoning Disruption Attack of LLMs (2025.acl-long)

Copied to clipboard

Challenge: Existing attacks on LLM reasoning are constrained by specific settings or lack of imperceptibility, limiting their feasibility and generalizability.
Approach: They propose a stepwise rEasoning error disruption attack that subtly injects errors into prior reasoning steps to mislead the model into producing incorrect subsequent reasoning and final answers.
Outcome: The proposed attack is compatible with zero-shot and few-shot settings, maintains the natural reasoning flow, and ensures covert execution without modification of the instruction.
Logic Jailbreak: Efficiently Unlocking LLM Safety Restrictions Through Formal Logical Expression (2026.findings-acl)

Copied to clipboard

Challenge: Despite advances in aligning LLMs with human values, current safety mechanisms remain vulnerable to jailbreak attacks.
Approach: They propose a black-box jailbreak method that uses logical expression translation to bypass LLM safety mechanisms.
Outcome: The proposed method exploits the distributional gap between alignment data and logic-expressed inputs while preserving the underlying semantic intent and readability while evading safety constraints.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations