森下 皓文 Terufumi Morishita
日立製作所・中央研究所の研究者です。 自然言語処理を主たる研究分野とし、機械学習や経済学×AIなどの周辺分野にも取り組んでいます：
- 自然言語処理： LLMの推論能力を向上させるための論理推論コーパスの自動生成
- 機械学習： アンサンブル学習の性能を決定づける根本要因の理論的な解明
- 経済学x AI： 多数のLLMエージェントを用いたマクロ経済シミュレーション
会社の実用寄りの研究の傍らで、このような、より基礎的な研究を進めています。
大学では、物理学(素粒子・宇宙物理学)を研究していました。
「現象の背後に潜む普遍的な原理」の探求に興味があります。 自然界の原理→物理学、思考の原理→記号論理学(論理推論)、経済現象の原理→経済学、というようにです。
I am a researcher at Hitachi’s Central Research Laboratory.
My primary research area is natural language processing, while I also work on adjacent areas such as machine learning and economics × AI:
- Natural language processing: automatic generation of logical-reasoning corpora for improving the reasoning capabilities of LLMs
- Machine learning: theoretical analysis of the fundamental factors that determine ensemble performance
- Economics × AI: large-scale macroeconomic simulation with LLM agents
Alongside practical, industry-oriented research, I pursue more fundamental research themes.
At university, I worked on physics, especially particle physics and astroparticle physics.
I am interested in uncovering the universal principles underlying phenomena. That includes the principles of nature in physics, the principles of thought in symbolic logic and logical reasoning, and the principles underlying economic phenomena in economics.
主な研究業績Selected Research Achievements
論文誌
国際学会
-
NeurIPS
Enhancing Reasoning Capabilities of LLMs via Principled Synthetic Logic Corpus link
Annual Conference on Neural Information Processing Systems
-
ICML
Learning Deductive Reasoning from Synthetic Corpus based on Formal Logic link
International Conference on Machine Learning
-
ICML
(Spotlight=上位5%)Rethinking Fano's Inequality in Ensemble Learning link
International Conference on Machine Learning
-
SemEval
(1位)Hitachi at SemEval-2020 Task 3: Exploring the representation spaces of transformers for human sense word similarity link
Proceedings of the Fourteenth Workshop on Semantic Evaluation
-
SemEval
(1位, Oral)Hitachi at SemEval-2020 task 7: Stacking at scale with heterogeneous language models for humor recognition link
Proceedings of the Fourteenth Workshop on Semantic Evaluation
-
SemEval
(2位)Hitachi at SemEval-2020 task 8: Simple but effective modality ensemble for meme emotion recognition link
Proceedings of the Fourteenth Workshop on Semantic Evaluation
国内学会
-
人工知能学会
EconTradeAgent: LLMエージェントと国際貿易理論に基づくマクロ経済シミュレーション
人工知能学会全国大会論文集 第40回
招待講演
-
AAAI
Can We Teach Logical Reasoning to LLMs? – An Approach Using Synthetic Corpora link
AAAI Bridge Program "Logical and Symbolic Reasoning in Language Models" keynote speech
International Conferences
-
NeurIPS
Enhancing Reasoning Capabilities of LLMs via Principled Synthetic Logic Corpus link
Annual Conference on Neural Information Processing Systems
-
ICML
Learning Deductive Reasoning from Synthetic Corpus based on Formal Logic link
International Conference on Machine Learning
-
ICML
Rethinking Fano's Inequality in Ensemble Learning link
International Conference on Machine Learning
-
SemEval
Hitachi at SemEval-2020 Task 3: Exploring the representation spaces of transformers for human sense word similarity link
Proceedings of the Fourteenth Workshop on Semantic Evaluation
-
SemEval
Hitachi at SemEval-2020 task 7: Stacking at scale with heterogeneous language models for humor recognition link
Proceedings of the Fourteenth Workshop on Semantic Evaluation
-
SemEval
Hitachi at SemEval-2020 task 8: Simple but effective modality ensemble for meme emotion recognition link
Proceedings of the Fourteenth Workshop on Semantic Evaluation