Dan Roth · page 3
Papers
- A Distributional and Orthographic Aggregation Model for English Derivational Morphology (2018 · 19 citations)
The task is derived word generation: given a root word and a transformation (for example "someone who runs" from "run" by AGENT), produce the derived word ("runner"). - Evaluation of named entity coreference (2019 · 14 citations)
The paper argues that the usual coreference scores can mislead when the real goal is to collect every reference to a named entity, whether by name, pronoun or nominal, as search or information extraction requires. - Coreference Reasoning in Machine Reading Comprehension (2021 · 11 citations)
The paper argues that Quoref, a benchmark built to test whether machine reading comprehension (MRC) models resolve coreference, overstates how well they do. - Weaver: Interweaving SQL and LLM for Table Reasoning (2025 · 6 citations)
Answering questions over tables is hard when cells or surrounding paragraphs hold free text (or images). SQL cannot interpret such content semantically, while LLMs struggle with long inputs and with counting, arithmetic and temporal logic. Earlier hybrids either embed LLM calls inside SQL, which breaks down on… - MoNaCo: More Natural and Complex Questions for Reasoning Across Dozens of Documents (2026 · 5 citations)
LLM-based agents are increasingly used to look up information. However, few benchmarks test them on realistic questions that would take a person hours of research across many sources. MoNaCo fills this gap with 1,315 human-written questions. Crowd workers wrote them while imagining the needs of specific personas… - LAD-RAG: Layout-aware Dynamic RAG for Visually-Rich Document Understanding (2026 · 4 citations)
Retrieval-augmented generation over long, visually rich documents (slides, reports, papers) often fails on questions whose answers are spread across several pages. - ToolScope: Enhancing LLM Agent Tool Use through Tool Merging and Context-Aware Filtering (2026 · 4 citations)
LLM agents often choose the wrong external tool when a toolset contains many near-duplicate tools with similar names and descriptions. Large toolsets also overflow the model's input window. ToolScope (Oracle AI) addresses both problems without fine-tuning. - MT-OSC: Path for LLMs that Get Lost in Multi-Turn Conversation (2026 · 3 citations)
LLMs perform noticeably worse when a task's instructions and context are spread across several dialogue turns. - Named Person Coreference in English News (2018 · 3 citations)
The paper looks at a narrow slice of coreference called Named Person Coreference (NPC). The task is to find every mention of a person who is named somewhere in a document, whether the mention is the name itself, a pronoun or a descriptive noun phrase. Using OntoNotes with gold NER labels, the authors show that people… - Chart-RL: Generalized Chart Comprehension via Reinforcement Learning with Verifiable Rewards (2026 · 1 citation)
Vision-language models (VLMs) often fail on charts they have not seen before. Chart questions call for symbolic and quantitative multi-step reasoning, and supervised fine-tuning (SFT) on domain data tends to overfit and cause forgetting. The authors propose Chart-RL, which extends the open-source VLM-R1 platform with… - When Vision-Language Models Judge Without Seeing: Exposing Informativeness Bias (2026 · 1 citation)
The paper looks at how reliable vision-language models are when used as automatic judges in pairwise answer comparison. - On Reference (In-)Determinacy in Natural Language Inference (2025 · 1 citation)
Most NLI datasets are annotated under reference determinacy (RD): annotators assume the premise and hypothesis describe the same situation, entities and events. - Decomposed Prompting Does Not Fix Knowledge Gaps, But Helps Models Say “I Don't Know” (2026)
In closed-book question answering, LLMs often fail to recognize what they do not know and produce confident hallucinations. - SPENCE: A Syntactic Probe for Detecting Contamination in NL2SQL Benchmarks (2026)
Leaderboard scores of LLMs on NL2SQL benchmarks may be overstated if models saw benchmark items, or close variants, during training.
This page in Markdown → https://md.research.elamur.ai/free/en/authors/dan-roth/page/3/