Parisa Rabbani, Priyam Sahoo, Ruben Mathew, Aishee Mondal, Harshita Ketharaman, Nimet Beyza Bozdag, Dilek Hakkani-Tür. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Large language models (LLMs) often exhibit sycophancy: agreement with user stance even when it conflicts with the model’s opinion. While prior work has mostly studied this in single-agent settings, it remains underexplored in collaborative multi-agent systems. We ask whether awareness of other agents’ sycophancy levels influences discussion outcomes. To investigate this, we run controlled experiments with six open-source LLMs, providing agents with peer sycophancy rankings that estimate each peer’s tendency toward sycophancy. These rankings are based on scores calculated using various static (pre-discussion) and dynamic (online) strategies. We find that providing sycophancy priors reduces the influence of sycophancy-prone peers, mitigates error-cascades, and improves final discussion accuracy by an absolute 10.5%. Thus, this is a lightweight and efficient way to reduce model sycophancy during discussions and subsequently improve downstream accuracy.

This paper describes a replication study of influential recent work on binary-level code stylometry by Caliskan et al. [8]. Using the Google Code Jam (GCJ) dataset that the original work used but with possible differences in authors and tasks, the accuracy results we obtain are significantly lower than those originally reported. An analysis of the features that contribute most to author classification decisions indicates that such features may, in many cases, be accidental artifacts---e.g., due to erroneous disassembly of data bytes embedded in the binary---and have little to do with programming style. Our results suggest that binary-level code stylometry. (1) is more sensitive to code characteristics than previously suspected; (2) can be significantly less accurate than previously reported (for 100 authors, we achieved approximately 63% accuracy, compared to the 96% reported in the original work); and (3) deserves careful attention to accidental artifacts arising from the compilation and stylometry toolchains. We found 29/33 of top ndisasm-based features resulted from erroneous disassembly. Our analysis revealed that this might cause the model to pick spurious features, i.e., the original file name, as the g++ compiler embeds the filename of the source CPP file into the binary -- which might unknowingly inflate the results.