Min Zhang · page 2
Papers
- MTSA: Multi-turn Safety Alignment for LLMs through Multi-round Red-teaming (2025 · 14 citations)
Jailbreak research and safety training have mostly targeted single-turn prompts. In multi-turn chats, an attacker can spread harmful intent across several innocent-looking turns, and LLMs are more likely to give in. MTSA is a two-stage framework for multi-turn safety. - Constituency Parsing using LLMs (2023 · 14 citations)
The paper asks whether modern large language models can produce phrase-structure trees, a task where supervised parsers reach about 96% while unsupervised methods remain below 65%. - Safety Alignment via Constrained Knowledge Unlearning (2025 · 13 citations)
Aligned large language models still produce harmful content under jailbreak attacks. The authors argue this happens because standard defenses, such as continued safety training, teach models to refuse without removing the dangerous knowledge itself, and they can also erode general skills. They propose Constrained… - Backdoors in RLVR: Jailbreak Backdoors in LLMs From Verifiable Reward (2026 · 10 citations)
This paper presents what the authors describe as the first backdoor attack on Reinforcement Learning with Verifiable Rewards (RLVR), the training approach used to strengthen LLM reasoning in math and code. - Speaker-Aware Discourse Parsing on Multi-Party Dialogues (2022 · 10 citations)
This paper addresses discourse parsing for conversations with several participants. The task is to build an SDRT-style dependency structure that links utterances (EDUs) and labels each link with a relation such as QAP or Elab. Earlier neural parsers focus mainly on conversational context. At best, they add a binary… - GUI-explorer: Autonomous Exploration and Mining of Transition-aware Knowledge for GUI Agent (2025 · 9 citations)
Multimodal LLM agents that operate mobile apps often misread interface elements, such as a music-recognition icon in QQ Music, and their built-in knowledge becomes outdated as apps update. Fine-tuning for each app version is too costly. GUI-explorer is a training-free agent that addresses both problems in three steps: - Efficient Reasoning for LLMs through Speculative Chain-of-Thought (2026 · 6 citations)
Reasoning LLMs such as Deepseek-R1 are slow because large models produce very long chains of thought (CoT). - A Systematic Survey of Semantic Role Labeling in the Era of Pretrained Language Models (2025 · 6 citations)
This survey organizes about 25 years of semantic role labeling (SRL) research, from Gildea and Jurafsky (2000) to work in the LLM era. - Dynamic Sampling that Adapts: Self-Aware Iterative Data Persistent Optimization for Mathematical Reasoning (2026 · 4 citations)
Data selection for training math-reasoning LLMs usually depends on fixed difficulty labels or external scorers. - MathAgent: Adversarial Evolution of Constraint Graphs for Mathematical Reasoning Data Synthesis (2026 · 3 citations)
Synthetic data for mathematical reasoning is often built either by mutating human-curated seed problems, which limits diversity to the seeds' scope, or by prompting models from scratch, which tends toward repetitive, shallow problems (mode collapse). - SAT: Balancing Reasoning Accuracy and Efficiency with Stepwise Adaptive Thinking (2026 · 3 citations)
Large Reasoning Models often "overthink" and produce long chains of thought even for simple problems. Existing remedies are either coarse, such as prompting, routing, or token budgeting, or apply one uniform rule to every step, such as early stopping or suppressing reflection. Both kinds can cut away steps that… - ComfyUI-R1: Exploring Reasoning Models for Workflow Generation (2026 · 1 citation)
ComfyUI is a node-based platform for building image and video generation pipelines. Assembling a working workflow there takes real expertise, because the platform has thousands of community components with complex dependencies. Earlier automation attempts prompted proprietary LLMs or multi-agent setups to emit JSON… - IS-CoT: Breaking the Long-form Generation Collapse via Interleaved Structural Thinking (2026 · 1 citation)
The paper tackles a weakness of LLMs in open-ended long-form writing: once the requested length goes beyond roughly 2,000 words, both quality and adherence to the target length drop sharply. - When Is Thinking Enough? Early Exit via Sufficiency Assessment for Efficient Reasoning (2026 · 1 citation)
Large reasoning models tend to overthink: they keep re-verifying or trying other strategies after they already have a correct solution, which wastes tokens. - Beyond Unimodal Shortcuts: MLLMs as Cross-Modal Reasoners for Grounded Named Entity Recognition (2026)
Grounded Multimodal Named Entity Recognition (GMNER) asks a system to find named entities in a sentence, label their types, and link each one to a bounding box in the paired image, or mark it as absent from the image. - Crossing the Reward Bridge: Expanding Reinforcement Learning with Verifiable Rewards Across Diverse Domains (2026)
RLVR works well for math and code, where a simple rule can check whether an answer is right. Such checkers fail on free-form answers, and the authors show that rule-based rewards can even get worse as more mixed-domain training data is added. The paper replaces the rule with a generative verifier. An LLM compares the… - Data Foundations of Long-Context Language Models: A Survey (2026)
Most earlier surveys of long-context language models (LCMs, here meaning models with windows above 128K tokens) focus on architecture and efficient training, so the role of data has been largely overlooked. - MSVBench: Towards Human-Level Evaluation of Multi-Shot Video Generation (2026)
Video generators increasingly produce story-like videos made of many shots, but most evaluation benchmarks still judge single clips. - SR-RAG: Verifiable Multi-Hop Reasoning via On-the-fly Symbolic Graph Construction (2026)
Multi-hop question answering is still difficult for retrieval-augmented generation. Aggregate-first systems such as GraphRAG build a corpus-wide knowledge graph in advance, which is costly and hard to keep current. Dynamic-first agents such as ReAct alternate retrieval and generation, but one wrong intermediate… - Think Faster Than Words: Efficient LLM Chain-of-Thought Reasoning via Dynamic Shortcut Decoding (2026)
Reasoning-oriented LLMs often keep producing chain-of-thought text after the problem is effectively solved.
This page in Markdown → https://md.research.elamur.ai/free/en/authors/min-zhang/page/2/