On the Advance of Making Language Models Better Reasoners

By Yifei Li, Zeqi Lin, Shizhuo Dylan Zhang, Qiang Fu, Jian-Guang Lou, Weizhu Chen. 2022

162 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Large Language Models, Prompting & Reasoning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Large Language Models

In Prompting & Reasoning

Citations in the corpus

Reasoning Like Program Executors (2022, background): “, GPT-3) can perform well on reasoning tasks (Wei et al., 2022; Kojima et al., 2022; Li et al., 2022b).”
Large Language Models Can Self-Improve (2022, neutral): “We end with ablation studies on model sizes and hyperpa-rameters.”
Rationale-Augmented Ensembles in Language Models (2022, background): “…does not require additional ﬁne-tuning (Thoppilan et al., 2022; Zelikman et al., 2022), veriﬁers (Cobbe et al., 2021), calibrators (Ye & Durrett, 2022), or any use of an auxiliary dataset (Zelikman et al., 2022; Li et al., 2022), making it applicable to any off-the-shelf large language model.”
PAL: Program-aided Language Models (2022, background): “Another line of work attempts to improve on chain-of-thought prompting by generating multiple reasoning chains (Li et al., 2022; Wang et al., 2022c) and 0-shot chain-of-thought prompting Kojima et al.”
Multimodal Chain-of-Thought Reasoning in Language Models (2023, neutral): “This section reviews recent progress of eliciting CoT reasoning by prompting and ﬁne-tuning language models.”
CodeT: Code Generation with Generated Tests (2022, background): “The idea of ranking based on agreement also appears in the domain of reasoning (Wang et al., 2022; Li et al., 2022a).”
Let's Verify Step by Step (2023, methodology): “We experimented with using RM-weighted voting (Li et al., 2022; Uesato et al., 2022) to combine the benefits of the PRM and majority voting, but this did not noticeably improve performance.”
Towards Reasoning in Large Language Models: A Survey (2022, background): “In addition, there are benchmarks for evaluating LLMs’ generative relational reasoning abilities, such as CommonGen (Lin et al., 2020; Liu et al., 2022a) and Open Relation Modeling (Huang et al., 2022b,d).”
Reasoning with Language Model Prompting: A Survey (2022, background): “, 2022g), as well as the continuity between reasoning steps (Li et al., 2022e) both should affect the accuracy of final answers.”
WizardMath: Empowering Mathematical Reasoning for Large Language Models via Reinforced Evol-Instruct (2023, unknown): “Historically, outcome-based approaches focused on algorithmic tasks (Li et al., 2016; Cai et al., 2017; Yu et al., 2023a), while recent research demonstrates the efficacy of reward models or validators in enhancing model performance (Cobbe et al., 2021; Wang et al., 2023a;b; Li et al., 2022).”
Unleashing the potential of prompt engineering in Large Language Models: a comprehensive review (2023, methodology): “[128] have introduced a novel prompting approach called Diverse Verifier on Reasoning Step (DIVERSE).”
Solving math word problems with process- and outcome-based feedback (2022, unknown): “The ORM-RL and PRM-RL models achieve a ﬁnal-answer error rate below 13%, improving on the 16.8% ﬁnal-answer error for the current state-of-the-art model (Li et al., 2022).”
Decomposition Enhances Reasoning via Self-Evaluation Guided Decoding (2023, unknown): “For instance, Wang et al. (2022b) introduce self-consistency as a method to determine the final answer through majority voting using multiple sampled reasoning paths, while Li et al. (2022) investigate various prompts to diversify the sampling outcomes.”
A Survey of Large Language Models (2023, unknown): “In contrast, other studies train a scoring model to measure the reliability of the generated reasoning paths [249] or continually train LLMs”
Self-Polish: Enhance Reasoning in Large Language Models via Problem Refinement (2023, background): “Another line of work optimizes the rationales for better consistency and continuity (Wang et al., 2022; Li et al., 2022; Zelikman et al., 2022; Zheng et al., 2023), and a representative one is Self-Consistency (SC).”
A Survey of Deep Learning for Mathematical Reasoning (2022, unknown): “Outcome-based approaches acknowledge the potential incorrectness of an individual reasoning path, and instead use multiple reasoning paths (Wang et al., 2022; Li et al., 2022a).”
L2CEval: Evaluating Language-to-Code Generation Capabilities of Large Language Models (2023, methodology): “Note that with a well-calibrated model, methods such as voting (Li et al., 2022a; Wang et al., 2022a) and confidencebased reranking (Ni et al.”
ChatGPT: Jack of all trades, master of none (2023, unknown): “7%) 1000 - 11 MathQA S EN Question answering No Mathematical reasoning GSM8K [62] / [63] 3 Yes 1319 1000 0 1 (0.”