Caiming Xiong · page 2
Papers
- DialFact: A Benchmark for Fact-Checking in Dialogue (2022 · 97 citations)
Automatic fact-checking has mostly focused on formal text such as news, Wikipedia sentences or tables. Conversational utterances are harder: they are informal, mix opinions with facts, and are full of ellipsis and coreference, yet no public benchmark existed for them. The authors frame dialogue fact-checking as three… - Find or Classify? Dual Strategy for Slot-Value Predictions on Multi-Domain Dialog State Tracking (2020 · 79 citations)
Dialog state tracking (DST) in task-oriented assistants must infer, turn by turn, every <domain, slot, value> triplet a user wants. - Summary of a Haystack: A Challenge to Long-Context LLMs and RAG Systems (2024 · 69 citations)
Long-context LLMs and RAG pipelines can now take in enormous inputs, but simple probes such as Needle-in-a-Haystack are too easy to separate strong systems. - Direct Judgement Preference Optimization (2025 · 35 citations)
LLM-as-judge models are usually trained with supervised finetuning on fairly small datasets and cover only narrow evaluation formats, often just pairwise comparison. - CRMArena: Understanding the Capacity of LLM Agents to Perform Professional CRM Tasks in Realistic Environments (2025 · 30 citations)
Testing LLM agents on real enterprise work is hard. Company CRM data is private, and existing work benchmarks such as WorkArena, WorkBench and Tau-Bench use simple data structures and easy tasks. The authors built CRMArena, a benchmark set inside a real Salesforce organization. With help from CRM practitioners, they… - L2CEval: Evaluating Language-to-Code Generation Capabilities of Large Language Models (2024 · 27 citations)
Large language models trained on code can turn natural-language requests into executable programs, but earlier evaluations covered isolated tasks, single model families or particular learning setups. - CodeTree: Agent-guided Tree Search for Code Generation with Large Language Models (2025 · 21 citations)
Hard programming tasks have a huge space of possible solutions. Older methods either sample many independent candidates or repeatedly refine one early attempt, which can trap them in a local optimum. CodeTree addresses this by organizing code generation as a tree search run by cooperating LLM agents. A Thinker agent… - DialogStudio: Towards Richest and Most Diverse Unified Dataset Collection for Conversational AI (2023 · 21 citations)
Conversational models often struggle across varied dialogue tasks because training data is fragmented, small, task-specific and stored in inconsistent formats. - Shared Imagination: LLMs Hallucinate Alike (2024 · 7 citations)
Modern LLMs are trained with very similar recipes, so the authors ask whether the resulting models are also similar in what they invent. - Simple Data Augmentation with the Mask Token Improves Domain Adaptation for Dialog Act Tagging (2020 · 5 citations)
Dialog act (DA) taggers learn speaker intentions such as "request" or "inform". These intentions are shared across task-oriented domains, yet taggers trained on one domain transfer poorly to another, and labeling new domains is costly. The authors adapt DA taggers using only unlabeled target-domain dialogues. Their… - J4R: Learning to Judge with Equivalent Initial State Group Relative Policy Optimization (2026 · 3 citations)
LLM-as-judge models handle simple tasks like chat quality well, but they struggle to evaluate reasoning-heavy outputs. - Hard2Verify: A Step-Level Verification Benchmark for Open-Ended Frontier Math (2026 · 2 citations)
Frontier LLMs now write Olympiad-level proofs, such as the gold-medal results at IMO 2025. Training and evaluating such systems requires automatic verifiers that can judge each step of a proof, not just the final answer. Existing step-level benchmarks have weaknesses: they use easy questions, outputs from weak… - From Passive Metric to Active Signal: The Evolving Role of Uncertainty Quantification in Large Language Models (2026 · 1 citation)
Most uncertainty quantification (UQ) work on LLMs scores an output only after it has been generated. This survey argues that such retrospective scoring does not fit multi-step reasoning, agents or alignment loops. It traces a shift in which uncertainty becomes a live control input that changes what the model does…
This page in Markdown → https://md.research.elamur.ai/free/en/authors/caiming-xiong/page/2/