Papers by Guotong Geng

7 papers
IS-CoT: Breaking the Long-form Generation Collapse via Interleaved Structural Thinking (2026.acl-long)

Copied to clipboard

Challenge: Existing models with reasoning capabilities suffer from a severe length collapse in open-ended writing .
Approach: They propose a framework that embeds a dynamic plan-write-reflect cycle into the generation process and train a model with interleaved reasoning traces.
Outcome: The proposed framework achieves state-of-the-art performance on long-form benchmarks compared to other models on the same dataset.
DisCal: Distribution-Aware Calibration for Mathematical Reasoning Under Character-Level Noisy Inputs (2026.acl-long)

Copied to clipboard

Challenge: Existing methods for calibration of large reasoning models (LRMs) focus on clean inputs, leaving noise unexplored.
Approach: They propose a confidence calibration framework for character-level noisy inputs that extracts uncertainty signals from both the empirical answer distribution and the model’s predictive distribution and integrates them via a learned calibrator.
Outcome: Experiments on multiple mathematical reasoning benchmarks show that DisCal outperforms existing calibration methods under noisy inputs, reducing expected calibration error (ECE) by up to 39.21% and improving Area Under the Receiver Operating Characteristic Curve (AUROC) by 31.44%.
SHARP: Self-adaptive Harmful Category-aware Prompt Generation for Black-box Jailbreaking (2026.acl-long)

Copied to clipboard

Challenge: Existing methods for jailbreak ignore the semantic differences between categories of harmful questions, leading to inconsistent success rates and reduced overall attack effectiveness.
Approach: They propose a category-aware jailbreak framework that incorporates the semantic category of harmful questions into prompt generation.
Outcome: The proposed framework improves attack success rates and category alignment and achieves better cross-category robustness compared to the state-of-the-art (SOTA) baselines.
A Simple Hash-Based Early Exiting Approach For Language Understanding and Generation (2022.findings-acl)

Copied to clipboard

Challenge: Existing methods to measure instance difficulty use generalization and threshold-tuning . a new approach to learn to exit is based on hash functions to assign tokens to a fixed exiting layer.
Approach: They propose a Hash-based Early Exiting approach that replaces learn-to-exit modules with hash functions to assign each token to a fixed exiting layer.
Outcome: The proposed approach improves on learning to exit and predicting instance difficulty.
When Rules Learn: A Self-Evolving Agent for Legal Case Retrieval (2026.findings-acl)

Copied to clipboard

Challenge: Existing dense retrieval methods have achieved notable progress, but their effectiveness in legal case retrieval remains limited.
Approach: They propose a self-evolving framework for rule-driven query rewriting that enhances BM25 without any parameter training.
Outcome: The proposed framework outperforms non-evolutionary baselines, including human-designed rules and greedy rule selection, especially when powered by a high-capacity core LLM.
SafeConf: A Confidence-Calibrated Safety Self-Evaluation Method for Large Language Models (2025.findings-emnlp)

Copied to clipboard

Challenge: Large language models (LLMs) have many advantages but they also pose significant safety risks.
Approach: They propose a method to enhance the safety self-evaluation capability of LLMs . they perform semantic mutations on the original safety evaluation questions .
Outcome: The proposed method improves safety self-evaluation accuracy by 5.86% and 7.79% over baseline methods on Chinese and English datasets.
Dynamic Evil Score-Guided Decoding: An Efficient Decoding Framework For Red-Team Model (2025.findings-acl)

Copied to clipboard

Challenge: Existing red-teaming methods require expensive fine-tuning, especially for large LLMs.
Approach: They propose a red-teaming method that uses an ‘evil score’ to evaluate the potential of tokens to contribute to harmful outputs during decoding.
Outcome: The proposed method achieves an ASR of 92.83% on the Llama-3.2-3B-Instruct model, compared to 83.48% with adversarial fine-tuning while using less computational resources.

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