Heng Ji · page 2
Papers
- CREATOR: Tool Creation for Disentangling Abstract and Concrete Reasoning of Large Language Models (2023 · 75 citations)
Tool-augmented LLMs usually depend on a fixed set of existing APIs. They must plan and execute in a single pass of implicit reasoning, which breaks easily, and they handle errors poorly. CREATOR addresses this by having the model write its own tools instead of only calling given ones. The framework has four stages.… - Using Semantic Relations to Refine Coreference Decisions (2005 · 67 citations)
Most coreference resolvers depend on local lexical and syntactic cues such as string matching, agreement and distance. - Massively Multi-Cultural Knowledge Acquisition & LM Benchmarking (2024 · 56 citations)
Large language models often reflect Western-centric biases and lack culture-specific commonsense, especially about variation within a country. - RM-R1: Reward Modeling as Reasoning (2025 · 49 citations)
Reward models guide the alignment of LLMs, but each existing family has a weakness. Scalar RMs give no justification for their scores. Generative RMs explain their verdicts, yet their reasoning is often shallow. The authors treat preference judgment as a reasoning problem and introduce Reasoning Reward Models… - Word Embeddings Are Steers for Language Models (2024 · 49 citations)
The paper asks what role output word embeddings play in how language models generate text, a question earlier work has mostly treated only at the level of individual word meaning. - A Context-Dependent Gated Module for Incorporating Symbolic Semantics into Event Coreference Resolution (2021 · 48 citations)
The paper addresses within-document event coreference resolution: grouping event mentions in a text that describe the same real-world event. - Persona-DB: Efficient Large Language Model Personalization for Response Prediction with Collaborative Data Refinement (2025 · 29 citations)
The paper addresses context efficiency in retrieval-augmented LLM personalization: how to keep predictions of a user's reactions accurate while placing fewer retrieved items into the prompt. - Aligning LLMs with Individual Preferences via Interaction (2025 · 24 citations)
Most LLM alignment work targets broad norms such as helpfulness, harmlessness and honesty, and ignores that individual users have different, often unstated, preferences. - Understanding Emergent In-Context Learning from a Kernel Regression Perspective (2023 · 21 citations)
Large language models can solve new tasks from a few demonstrations in the prompt (in-context learning, ICL) without any weight updates, but the mechanism behind this ability remains unclear. - Harnessing the Reasoning Economy: A Survey of Efficient Reasoning for Large Language Models (2025 · 16 citations)
Large Reasoning Models such as OpenAI o1, DeepSeek R1 and QwQ solve hard math and coding problems by producing long chains of thought. - Current Agents Fail to Leverage World Model as Tool for Foresight (2026 · 1 citation)
The paper asks whether today's vision-language agents can treat a generative world model as an optional external tool: a simulator they consult to preview the consequences of an action before committing to it in the real environment. - Beyond Sample-Level Feedback: Using Reference-Level Feedback to Guide Data Synthesis (2026)
Synthetic instruction-tuning data is cheap to produce, but a model trained on it usually cannot exceed the LLM that generated it.
This page in Markdown → https://md.research.elamur.ai/free/en/authors/heng-ji/page/2/