RIKEN AIP Researchers Land 9 Papers in EMNLP 2026 Main Conference

Acceptance figures

RIKEN AIP describes EMNLP, the Conference on Empirical Methods in Natural Language Processing, as one of the leading international conferences in natural language processing. The center says papers involving its researchers were accepted across several venues: nine to the Main Conference, seven to Findings of EMNLP 2026, and two to other tracks and co-located conferences.

The two remaining papers are listed under the System Demonstration Track and under WMT 2026, the Conference on Machine Translation. Many of the papers are joint work with other institutions, including Tohoku University, Kyoto University, The University of Tokyo, MBZUAI, the University of Sheffield and LMU Munich.

Main Conference papers

The nine Main Conference papers cover a range of topics in language model research. Titles listed by RIKEN AIP include:

• "A Robust Evaluation of Probe Robustness: Lessons for Reliable OOD Uncertainty Quantification"
• "It's Not RoPE That Creates Sinks: The Role of Self-Concentration and Value-Non-Mixing in Attention"
• "Weakening Neurons: An Input-Output Functionality in Transformers with Outsize Influence"
• "When Queries Crowd: Diagnosing and Improving Spread-Out Regularization for Text Embeddings"
• "Length and Meaning for Language Model Maps"
• "Language Model Maps for Prompt-Response Distributions via Log-Likelihood Vectors"
• "Reasoning Depth and Environment Complexity: A Controlled Study of RLVR Data Allocation across Logical Reasoning Tasks"
• "Cross-Lingual Representation Alignment by Token-Level Optimal Transport in a Language-Agnostic Space"
• "(How) Do MLLMs Report Bistable Images Like Humans?"

Authors from RIKEN-affiliated groups appear on all of them. Kentaro Inui and Benjamin Heinzerling are among the authors of the probe robustness paper, while Hidetoshi Shimodaira is listed on the language model map papers and the RLVR data allocation study. Hitomi Yanaka appears on the cross-lingual alignment paper and the bistable images paper.

Findings, demo and translation papers

The seven Findings papers include work on Arabic dialect steering in LLMs, author-personalized text generation, positional recall in Mamba language models, and domain mixture design for aligning language models with a target model. Two of them deal with security: one on neutral prompting attacks that steer hallucinations in agent skills, and one on a single suffix for jailbreaking merged model families. A further Findings paper covers compression of mixture-of-experts models through a shared low-rank basis factorization.

In the System Demonstration Track, the listed paper is "CAT: A CounterArgument Tutoring System based on Logical Structure Analysis". At WMT 2026, the listed paper is "Doc2FRC: Length-Consistent Document-Level Machine Translation via Fixed-Range Chunking".

What it means

The announcement is a list of accepted papers and does not report results or evaluation details for any of them. It does show the breadth of work that RIKEN AIP researchers and their partners are contributing to NLP, from analysis of transformer internals and attention behavior to multilingual representation, machine translation, and the security of language models. Readers who want details on individual papers can consult EMNLP 2026, to which RIKEN AIP points for more information.