Percy Liang · page 2
Papers
- Contrastive Decoding: Open-ended Text Generation as Optimization (2022 · 253 citations)
Open-ended text generation with large language models faces a tradeoff. Picking the most probable continuation tends to yield short, looping text, while sampling often wanders off topic or loses coherence. The authors introduce contrastive decoding (CD), a training-free, search-based method that pairs two frozen… - Decoupling Strategy and Generation in Negotiation Dialogues (2018 · 236 citations)
End-to-end neural negotiation agents learn strategy and wording together. This makes their behaviour hard to steer or inspect, and reinforcement learning tends to collapse their language into repetitive or ungrammatical output. The authors separate the two concerns with a three-part pipeline: - From Language to Programs: Bridging Reinforcement Learning and Maximum Marginal Likelihood (2017 · 236 citations)
The paper tackles weakly supervised semantic parsing, where natural language instructions must be turned into executable programs but training data provide only the correct execution outcome, not the program itself. - Demonstrate-Search-Predict: Composing retrieval and language models for knowledge-intensive NLP (2022 · 222 citations)
Pairing a frozen language model with a retriever usually follows a fixed recipe: fetch passages for the query, insert them into the prompt, and let the model answer. - Enabling Language Models to Fill in the Blanks (2020 · 209 citations)
Standard left-to-right language models generate text well but can only use preceding context, so they cannot fill gaps in the middle of a document. - MLAgentBench: Evaluating Language Agents on Machine Learning Experimentation (2023 · 147 citations)
MLAgentBench asks whether agents built on large language models can run machine learning experiments on their own: design an approach, write and execute code, read the results, and iterate. - Truncation Sampling as Language Model Desmoothing (2022 · 95 citations)
Sampling directly from neural language models often yields long texts that drift into nonsense. Truncation methods such as top-k and top-p help by zeroing out unlikely tokens, but it has been unclear what they are actually trying to achieve. The authors treat any neural LM as a blend of the true distribution with a… - Imitation Learning of Agenda-based Semantic Parsers (2015 · 95 citations)
Semantic parsers that answer questions over large knowledge bases like Freebase usually build logical forms bottom-up, span by span, keeping a fixed-size beam per chart cell. - BioMedLM: A 2.7B Parameter Language Model Trained On Biomedical Text (2024 · 79 citations)
Frontier systems such as GPT-4 and Med-PaLM 2 perform very well on biomedical NLP. However, they are huge, expensive, reachable only through remote APIs (a problem for patient privacy) and trained on undisclosed data. The authors ask whether a much smaller, domain-focused model can be competitive. They built… - Macro Grammars and Holistic Triggering for Efficient Semantic Parsing (2017 · 67 citations)
Training a semantic parser from question–answer pairs alone is expensive. Without annotated logical forms, the learner must explore a huge space of candidate programs to find ones that produce the correct answer. Zhang, Pasupat and Liang address this with an online learning algorithm whose search gets cheaper over… - RNNs can generate bounded hierarchical languages with optimal memory (2020 · 65 citations)
RNN language models capture natural-language syntax well in practice, but the theory behind this has been thin. - Planning, Inference and Pragmatics in Sequential Language Games (2018 · 31 citations)
The paper models how people talk in cooperative games where each partner holds hidden information. To win, a player must reason about three things. It must guess what the partner knows from their messages (inference). It must pick messages with future turns in mind (planning). It must account for the fact that the… - The Mighty ToRR: A Benchmark for Table Reasoning and Robustness in LLMs (2026 · 1 citation)
Large language models are often given tables, yet it is unclear how well they reason over them and how much their answers depend on how a table is written out.
This page in Markdown → https://md.research.elamur.ai/free/en/authors/percy-liang/page/2/