--- title: "Percy Liang — papers, page 2 — Elamur Research" description: "Percy Liang, Cornell University: 51 papers on Elamur Research (2006–2026), h-index 85 — affiliations, topics, co-authors and who cites them…" canonical: "https://research.elamur.ai/free/en/authors/percy-liang/page/2/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/authors/percy-liang/page/2/) [home](https://md.research.elamur.ai/free/en/) / [authors](https://md.research.elamur.ai/free/en/authors/) / [Percy Liang](https://md.research.elamur.ai/free/en/authors/percy-liang/) / page 2 # Percy Liang · page 2 51 papers in our library · 2006–2026 · page 2 of 3 [The profile, affiliations and co-authors →](https://md.research.elamur.ai/free/en/authors/percy-liang/) ## Papers The most cited first (OpenAlex and our library). [← prev](https://md.research.elamur.ai/free/en/authors/percy-liang/) [1](https://md.research.elamur.ai/free/en/authors/percy-liang/) **2** [3](https://md.research.elamur.ai/free/en/authors/percy-liang/page/3/) [next →](https://md.research.elamur.ai/free/en/authors/percy-liang/page/3/) 26. **[An Explanation of In-context Learning as Implicit Bayesian Inference](https://md.research.elamur.ai/free/en/papers/xie-2021-explanation-context-learning-implicit-bayesian/)** *(2021 · 286 citations)* The paper asks why large LMs such as GPT-3 can pick up a task from input-output examples in a prompt, even though they were never trained to learn from examples. 27. **[A Simple Domain-Independent Probabilistic Approach to Generation](https://md.research.elamur.ai/free/en/papers/angeli-2010-simple-domain-independent-probabilistic-approach/)** *(2010 · 264 citations)* The paper addresses data-to-text generation: producing a natural-language description from a world state given as a set of database records. 28. **[Contrastive Decoding: Open-ended Text Generation as Optimization](https://md.research.elamur.ai/free/en/papers/li-2022-contrastive-decoding-open-ended-text/)** *(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… 29. **[Decoupling Strategy and Generation in Negotiation Dialogues](https://md.research.elamur.ai/free/en/papers/he-2018-decoupling-strategy-generation-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: 30. **[From Language to Programs: Bridging Reinforcement Learning and Maximum Marginal Likelihood](https://md.research.elamur.ai/free/en/papers/guu-2017-language-programs-bridging-reinforcement-learning/)** *(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. 31. **[Learning a SAT Solver from Single-Bit Supervision](https://md.research.elamur.ai/free/en/papers/selsam-2018-learning-sat-solver-single-bit/)** *(2018 · 232 citations)* Selsam et al. ask whether a neural network can learn to solve propositional satisfiability (SAT), an NP-complete problem, when it is told only whether each instance is satisfiable. They build NeuroSAT, a message passing neural network (MPNN). It represents a CNF formula as a graph of literal and clause nodes and… 32. **[Demonstrate-Search-Predict: Composing retrieval and language models for knowledge-intensive NLP](https://md.research.elamur.ai/free/en/papers/khattab-2022-demonstrate-search-predict-composing-retrieval/)** *(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. 33. **[Transforming Question Answering Datasets Into Natural Language Inference Datasets](https://md.research.elamur.ai/free/en/papers/demszky-2018-transforming-question-answering-datasets-natural/)** *(2018 · 220 citations)* Widely used NLI benchmarks such as SNLI and MultiNLI cover phenomena like negation and quantification but rarely require reasoning across several sentences. 34. **[What Can Transformers Learn In-Context? A Case Study of Simple Function Classes](https://md.research.elamur.ai/free/en/papers/garg-2022-transformers-learn-context-case-study/)** *(2022 · 209 citations)* GPT-3 and similar models can solve tasks from examples in a prompt, but it is unclear whether they learn new tasks this way or only recall tasks from pretraining. 35. **[Enabling Language Models to Fill in the Blanks](https://md.research.elamur.ai/free/en/papers/donahue-2020-enabling-language-models-fill-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. 36. **[MLAgentBench: Evaluating Language Agents on Machine Learning Experimentation](https://md.research.elamur.ai/free/en/papers/huang-2023-mlagentbench-evaluating-language-agents/)** *(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. 37. **[Simpler Context-Dependent Logical Forms via Model Projections](https://md.research.elamur.ai/free/en/papers/long-2016-simpler-context-dependent-logical-forms/)** *(2016 · 145 citations)* The paper tackles learning to carry out multi-step instructions when the only supervision is the resulting world state, not the logical forms behind it. 38. **[Unifying Human and Statistical Evaluation for Natural Language Generation](https://md.research.elamur.ai/free/en/papers/hashimoto-2019-unifying-human-statistical-evaluation-natural/)** *(2019 · 143 citations)* Evaluating text generators is hard because the usual tools each miss something. Human ratings reward fluent output but cannot tell when a model just copies training data or always produces the same answer. Perplexity rewards coverage but barely punishes occasional bad samples. Hashimoto, Zhang and Liang define one… 39. **[LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals](https://md.research.elamur.ai/free/en/papers/park-2024-llm-agents-grounded-self-reports/)** *(2024 · 119 citations)* Machine-learning models that predict human behavior usually need large amounts of structured data for each specific outcome, so they are hard to reuse for new questions. 40. **[Reinforcement Learning on Web Interfaces Using Workflow-Guided Exploration](https://md.research.elamur.ai/free/en/papers/liu-2018-reinforcement-learning-web-interfaces-using/)** *(2018 · 119 citations)* The paper tackles a core obstacle in teaching RL agents to operate websites (for example, booking a flight or forwarding an email). 41. **[Robust Distortion-free Watermarks for Language Models](https://md.research.elamur.ai/free/en/papers/kuditipudi-2023-robust-distortion-free-watermarks-language/)** *(2023 · 110 citations)* The paper tackles provenance of LM-generated text. It seeks a watermark that leaves the model's output distribution unchanged (distortion-free), can be verified without access to the model or prompt (agnostic), and survives editing (robust). Earlier hashing-based schemes (Kirchenbauer et al., Aaronson, Christ et al.)… 42. **[Retrieval-Augmented Multimodal Language Modeling](https://md.research.elamur.ai/free/en/papers/yasunaga-2022-retrieval-augmented-multimodal-language-modeling/)** *(2022 · 104 citations)* Large multimodal generators such as DALL-E and CM3 keep all of their world knowledge inside their weights. 43. **[SPoC: Search-based Pseudocode to Code](https://md.research.elamur.ai/free/en/papers/kulal-2019-spoc-search-based-pseudocode-code/)** *(2019 · 96 citations)* SPoC addresses the generation of complete, multi-line C++ programs (roughly 10–20 lines) from line-by-line natural-language pseudocode. 44. **[Truncation Sampling as Language Model Desmoothing](https://md.research.elamur.ai/free/en/papers/hewitt-2022-truncation-sampling-as-language-model/)** *(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… 45. **[Imitation Learning of Agenda-based Semantic Parsers](https://md.research.elamur.ai/free/en/papers/berant-2015-imitation-learning-agenda-based-semantic/)** *(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. 46. **[BioMedLM: A 2.7B Parameter Language Model Trained On Biomedical Text](https://md.research.elamur.ai/free/en/papers/bolton-2024-biomedlm-2-7b-parameter-language/)** *(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… 47. **[Large Language Models as Analogical Reasoners](https://md.research.elamur.ai/free/en/papers/yasunaga-2023-large-language-models-analogical-reasoners/)** *(2023 · 74 citations)* Chain-of-thought prompting helps LLMs solve multi-step problems, but both common forms have drawbacks. The zero-shot variant gives only generic guidance, and the few-shot variant needs hand-written reasoning demonstrations for every task. Yasunaga et al. (Google DeepMind and Stanford) propose analogical prompting. It… 48. **[Macro Grammars and Holistic Triggering for Efficient Semantic Parsing](https://md.research.elamur.ai/free/en/papers/zhang-2017-macro-grammars-holistic-triggering-efficient/)** *(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… 49. **[RNNs can generate bounded hierarchical languages with optimal memory](https://md.research.elamur.ai/free/en/papers/hewitt-2020-rnns-can-generate-bounded-hierarchical/)** *(2020 · 65 citations)* RNN language models capture natural-language syntax well in practice, but the theory behind this has been thin. 50. **[Planning, Inference and Pragmatics in Sequential Language Games](https://md.research.elamur.ai/free/en/papers/khani-2018-planning-inference-pragmatics-sequential-language/)** *(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… [← prev](https://md.research.elamur.ai/free/en/authors/percy-liang/) [1](https://md.research.elamur.ai/free/en/authors/percy-liang/) **2** [3](https://md.research.elamur.ai/free/en/authors/percy-liang/page/3/) [next →](https://md.research.elamur.ai/free/en/authors/percy-liang/page/3/)