Shared Tasks

1The 4th Multilingual and Multimodal Medical Instructional Video Understanding Evaluation: Difficulty-Aware Challenge

DA-MIVQA extends the medical instructional video question answering tasks held at NLPCC from 2023 to 2025 by introducing difficulty-aware evaluation. Rather than measuring only overall performance, the new benchmark explicitly distinguishes between simple questions, which can be answered primarily through direct subtitle-aligned evidence, and complex questions, which require the integration of subtitles, visual demonstrations, and procedural context. This design aims to better assess model robustness in realistic medical education and emergency knowledge-access scenarios.

2Schwartz's Basic Human Values' Detection and Alignment with LLMs

LLM alignment has advanced rapidly in recent years. With the large-scale deployment of large language models in areas such as education, healthcare, psychological counseling, and public services, fine-grained value alignment has become an important prerequisite for safe and reliable AI systems. Existing studies often focus on coarse-grained compliance and bottom-line risk control, while systematic research on psychological basic human values remains limited. This shared task provides a standardized dataset based on Schwartz's theory of 19 basic human values and existing psychological test questions to support research on fine-grained value detection and alignment in LLMs.

3The Multi-dimensional Evaluation of Traditional Chinese Medicine LLMs: Capability, Reliability, and Robustness

While large language models have shown strong performance on knowledge-intensive benchmarks, existing evaluations of Traditional Chinese Medicine (TCM) LLMs largely focus on core domain knowledge and overlook reliability and robustness under perturbed question formats. This shared task introduces a multi-dimensional evaluation framework built from authoritative undergraduate textbook exercises to jointly assess capability, reliability, and robustness through question answering, statement verification, and format-perturbed inputs. The task aims to move beyond accuracy-centric evaluation and provide a more faithful assessment of how well TCM knowledge is internalized by the models.

4LLM-based Investment Advisor Agents for Asset Allocation in the Chinese Market

This task evaluates the ability of LLM-based investment advisor agents to perform complex reasoning and quantitative decision-making in the Chinese capital market. Moving beyond traditional text analysis, the task challenges participants to develop agents that interpret daily macroeconomic signals and sectoral shifts in order to execute daily asset-allocation strategies. Operating in a backtesting environment, the agents are provided with a Top-20 Financial Hot News feed and historical price data, and must autonomously generate daily rebalancing instructions for designated ETF pools.

Website: https://github.com/splash-li/NLPCC2026-Shared-Task-4/

5Chinese BabyLM: Data-Efficient and Developmentally Plausible Language Models for Chinese

Large language models achieve remarkable performance by training on trillions of words, yet human children acquire robust linguistic competence from far less input. Inspired by the BabyLM Challenge, Chinese BabyLM is the first shared task dedicated to sample-efficient pretraining for Chinese. Participants train language models on a constrained corpus of approximately 100 million words of developmentally plausible Chinese text, including children’s storybooks, child-adult conversations, and movie subtitles.

Website: https://chinese-babylm.github.io/

6The 2nd Shared Task on LLM-Generated Text Detection

The rapid development of large language models has introduced serious challenges, including disinformation generation, harmful content dissemination, and misuse. Against this backdrop, efficiently distinguishing LLM-generated text from human-written text has become an urgent research problem. Following the success of the first shared task at NLPCC 2025, the 2026 edition extends the task from binary classification to ternary classification by distinguishing among human-written text, LLM-generated text, and LLM-refined text. This setting better reflects real-world use cases and aims to advance robust Chinese LLM-generated text detection.

Website: https://nlp2ct.github.io/NLPCC-2026-Task6-Detection/

7ELSSTBenchmark: A Joint Evaluation of Implicit Concept Retrieval and Knowledge Discovery for LLMs

This shared task evaluates large language models on identifying implicit social-science concepts from long English texts. Given a 600–900-word English document written in a “show, don’t tell” style, where target concepts are implied through situations and narratives rather than explicitly named, participating systems may enter one or both independent tracks.

8Factivity Inference Inconsistency Attack (FIIA)

Factivity Inference (FI) is an important semantic understanding task concerned with the truthfulness of events. Although accurate access to factual information is crucial for large language models, existing models often exhibit instability in FI when subjected to minor textual perturbations or complex contexts. This shared task adopts a red-teaming setting to systematically expose the boundaries and vulnerable scenarios of current LLMs in complex FI tasks. Participants are required to creatively adapt the provided Chinese FI dataset to induce hallucinations or trigger inconsistency in model judgments.

9AISB (AI Scientist Benchmark)

AISB (AI Scientist Benchmark) evaluates the capability of AI systems to conduct scientific research autonomously. Given a research topic, reference papers, and established benchmarks with known baselines, participating AI Scientist systems must independently read the literature, formulate hypotheses, design experiments, write and execute code, analyze results, and produce a complete research paper without human intervention.

Website: https://github.com/ResearAI/NLPCC-2026-Task9-AISB

10Reliability of AI-Assisted Scientific Reporting

As generative AI and agentic AI become increasingly integrated into scientific workflows, they are now widely used to assist with scientific writing, including summarizing experimental results, drafting conclusions, and generating citation-supported statements. However, recent studies have shown that AI-assisted scientific reporting often overgeneralizes conclusions beyond what the source evidence justifies. This shared task focuses on determining whether an AI-generated scientific statement faithfully reflects the evidence it summarizes or cites.

Website: https://nlp2ct.github.io/NLPCC-2026-Task10-Science/

11Agent-Based Experiment Reproduction from Scientific Papers

As LLM-based AI systems become increasingly integral to scientific workflows, they provide substantial support across multiple stages of research. However, experiment reproduction remains a significant challenge for current AI agents. This task evaluates agents’ ability to reconstruct experimental pipelines from research papers by interpreting text, planning implementations, generating code, and executing experiments within a controlled environment. Unlike prior work that focuses only on final outputs, this task emphasizes the full reproduction process through fine-grained action logs and predefined rubrics covering key stages such as paper understanding, planning, coding, and execution.