Jianfeng Gao · page 2
Papers
- Towards End-to-End Reinforcement Learning of Dialogue Agents for Information Access (2017 · 350 citations)
The paper addresses how dialogue agents that help users find an entity in a Knowledge Base (KB-InfoBots) can be trained end-to-end. - A Neural Network Approach to Context-Sensitive Generation of Conversational Responses (2015 · 329 citations)
Open-domain chatbots built from social media data, such as Ritter et al. (2011)'s translation-style system, usually treat a reply as a function of the previous post alone. They ignore earlier turns. Adding that history to phrase-based statistical models makes phrase tables grow and become sparse. The authors address… - Chameleon: Plug-and-Play Compositional Reasoning with Large Language Models (2023 · 327 citations)
Large language models struggle with recent information, exact arithmetic and logic, and calling specialized tools. - SMART: Robust and Efficient Fine-Tuning for Pre-trained Natural Language Models through Principled Regularized Optimization (2020 · 304 citations)
Fine-tuning large pre-trained language models on small downstream datasets often overfits, and the usual remedies are hand-tuned heuristics such as special learning-rate schedules or freezing layers. The authors propose SMART, a regularized optimization framework for fine-tuning with two components. - Few-shot Natural Language Generation for Task-Oriented Dialog (2020 · 256 citations)
Task-oriented dialog systems need an NLG component that turns a dialog act, meaning an intent plus slot-value pairs, into a fluent reply. - Understanding the Difficulty of Training Transformers (2020 · 231 citations)
Transformers are hard to train. They often need warmup, carefully tuned adaptive optimizers and learning-rate schedules, and the original Post-LN layout tends to diverge as depth grows. This paper looks for the cause, using both theory and experiments. The authors first show that gradient vanishing is not the main… - Optimus: Organizing Sentences via Pre-trained Modeling of a Latent Space (2020 · 208 citations)
Variational Autoencoders (VAEs) can learn compact sentence representations that allow controllable generation. - Using Contextual Speller Techniques and Language Modeling for ESL Error Correction (2008 · 204 citations)
The paper tackles automatic proofing for non-native (ESL) writers. It focuses on determiner and preposition errors, which standard grammar checkers built for native speakers mostly ignore. Annotated learner-error data was not available, so the authors trained on well-formed native text: Encarta and Reuters sentences.… - Model Tells You What to Discard: Adaptive KV Cache Compression for LLMs (2023 · 202 citations)
Large language models cache the key and value vectors of every previous token during generation. This KV cache grows with model size and output length and often exceeds GPU memory. Offloading it to CPU or NVMe slows inference, and most token-pruning methods require retraining. The authors propose FastGen, a… - Image-Grounded Conversations: Multimodal Context for Natural Question and Response Generation (2017 · 199 citations)
This paper introduces Image-Grounded Conversations (IGC), a task that sits between open-ended chit-chat and goal-oriented dialog. - Generation-Augmented Retrieval for Open-domain Question Answering (2020 · 151 citations)
Retrieval is the bottleneck of retriever-reader systems for open-domain question answering (OpenQA). Sparse methods like BM25 are cheap but miss passages that share no words with the question. Dense retrievers such as DPR match meaning better, but they are costly to train and index. Mao et al. propose… - End-to-End Task-Completion Neural Dialogue Systems (2017 · 146 citations)
Classic task-completion dialogue systems chain separately trained modules (LU, dialogue manager, NLG), so mistakes made early propagate and pile up downstream. - AdaMix: Mixture-of-Adaptations for Parameter-efficient Model Tuning (2022 · 135 citations)
Full fine-tuning of large pre-trained language models requires a separate multi-hundred-megabyte copy of the weights for every task. - GO FIGURE: A Meta Evaluation of Factuality in Summarization (2021 · 130 citations)
Neural summarizers often produce fluent text that misstates or invents facts. Many new automatic metrics aim to detect such errors, but there has been no systematic way to check whether these metrics actually work. GO FIGURE is a meta-evaluation framework that tests factuality metrics against five requirements… - How Much Do Language Models Copy From Their Training Data? Evaluating Linguistic Novelty in Text Generation Using RAVEN (2023 · 117 citations)
When researchers judge a language model by the text it produces, they usually assume the output reflects learned linguistic knowledge. - Rethinking Interpretability in the Era of Large Language Models (2024 · 87 citations)
This position paper from Microsoft Research asks how interpretable machine learning should evolve now that large language models can describe patterns in plain language. - Self-Checker: Plug-and-Play Modules for Fact-Checking with Large Language Models (2024 · 87 citations)
LLM outputs often contain hallucinated facts. Most automated fact-checking systems depend on costly fine-tuning, and existing benchmarks mostly test short claims taken from Wikipedia rather than long, information-dense LLM answers. The authors propose SELF-CHECKER, a training-free framework built from prompted LLM… - Language Models as Inductive Reasoners (2024 · 70 citations)
Classical computational induction, such as Inductive Logic Programming, works over symbolic first-order logic. - Self-Verification Improves Few-Shot Clinical Information Extraction (2023 · 51 citations)
Clinical information extraction often lacks labeled training data. Few-shot prompting of LLMs can stand in for supervised training, but its outputs are still error-prone and hard to inspect, which matters in a safety-critical field like medicine. The authors propose self-verification (SV), a four-call chain in which… - SynthAgent: Adapting Web Agents with Synthetic Supervision (2026)
Web agents often fail on websites they never saw during training. Such sites rarely come with tasks or demonstrations, and hiring people to write them for each new site is costly. Earlier synthetic-data pipelines fix only one half of the task–trajectory pair. Some produce hallucinated, infeasible goals. Others…
This page in Markdown → https://md.research.elamur.ai/free/en/authors/jianfeng-gao/page/2/