More than 20 papers from Meituan’s technical team have been accepted by EMNLP 2026, including 12 main-conference papers covering streaming speech, model evaluation, reinforcement learning and agent training. The shared direction of this research is clear: the industry is shifting its attention from “what can the model do” to “whether it can complete work reliably with low latency and low cost.”
What this is
These papers are not a product about to launch; they are foundational research into large models and agents—model systems that can autonomously plan and call tools to complete tasks.
Among them, LTS-Agent seeks to let voice assistants continue reasoning while the user is speaking, reducing the wait associated with the traditional “recognition—reasoning—synthesis” process. Latent-GRPO generates reward signals from a model’s internal representations, reducing the cost of relying on external verifiers for reinforcement learning. FlexBench/FlexEval dynamically generates tests from specific work instructions to determine whether a model has genuinely completed the task. SIRI distills successful experiences into model parameters, avoiding the need to invoke an external skill library on every call.
In experiments, SIRI used Qwen2.5-7B-Instruct to raise the ALFWorld task success rate from 90.8% to 93.0%, and the WebShop rate from 72.8% to 81.3%. These results, however, remain paper and controlled-experiment findings and cannot be treated as equivalent to reliable gains in production environments.
Industry view
We believe the important signal is not the number of papers but the change in research objectives: the bottleneck in enterprise AI is gradually moving away from model intelligence and toward system efficiency. Speech latency, training costs, evaluation methods and skill reuse in long-running tasks all directly determine whether an application is usable and economically viable.
The risks are equally clear. First, paper experiments typically rely on specific models, datasets and metrics; when transferred to real business operations, their benefits may be diluted by data quality, tool reliability and workflow complexity. Second, reducing external skill retrieval can lower latency but may make models harder to audit and update. When internal business rules change, knowledge internalized into parameters may not be easy to correct. FlexBench also shows that the same model can rank differently and fail in different ways under different objective instructions, making a single general-purpose leaderboard increasingly inadequate as a measure of real-world work capability.
Impact on regular people
For enterprise IT: When procuring AI systems, teams should spend less time looking at demo rankings and ask vendors more often about completion rates under specific instructions, response latency and failure-recovery mechanisms. Although technical approaches are becoming more complex, selection criteria should move closer to business outcomes.
For careers: People who use AI will gradually move beyond simply writing prompts toward decomposing objectives, validating results and designing workflows. Judgment and process-design skills will matter more than raw generation speed.
For the consumer market: Voice assistants are likely to become more natural at listening and responding simultaneously, but these advances still require validation before large-scale commercial deployment. For now, consumers choosing products should continue to prioritize response speed, accuracy and whether errors can be corrected.