Sudong Cai

Chuang Ma, Qianying Liu, Tomoyuki Obuchi et al.

Multimodal large language models (MLLMs) remain unreliable on spatial multiple-choice questions, and their failures are often attributed to poorly attended visual information. In this work, we identify a complementary failure mode, spatial lexical bias: adding a spatial relation word to the answer options can attract the model's decision and make the newly added option likely to be selected. Using nine open-weight MLLMs, we show that this phenomenon is widely observed. In particular, models can answer a binary spatial question correctly, yet consistently select an incorrect third spatial option once it is added to the answer set. We isolate such binary-stable but ternary-fragile cases as diagnostic examples and leverage mechanistic interpretability tools, revealing that a substantial part of the failure instead originates on the language side rather than the visual side: visual attention analyses and residual-stream probes show the correct spatial relation remains internally available on these failures, while irrelevant-option controls, activation patching, and sparse component interventions trace the bias to specific LLM-side channels and neurons. Based on this finding, we show that a lightweight LLM-only DPO update on tiny single-object-pair synthetic data mitigates the bias, lifting four-way robust accuracy by up to 100 points on synthetic data, and by 68.0, 32.6, and 20.1 points on broader evaluation datasets WhatsUp, SpatialMQA-Direct, and VSR.

Shuyuan Zheng, Sudong Cai, Chuan Xiao et al.

In collaborative machine learning (CML), data valuation, i.e., evaluating the contribution of each client's data to the machine learning model, has become a critical task for incentivizing and selecting positive data contributions. However, existing studies often assume that clients engage in data valuation truthfully, overlooking the practical motivation for clients to exaggerate their contributions. To unlock this threat, this paper introduces the data overvaluation attack, enabling strategic clients to have their data significantly overvalued in federated learning, a widely adopted paradigm for decentralized CML. Furthermore, we propose a Bayesian truthful data valuation metric, named Truth-Shapley. Truth-Shapley is the unique metric that guarantees some promising axioms for data valuation while ensuring that clients' optimal strategy is to perform truthful data valuation under certain conditions. Our experiments demonstrate the vulnerability of existing data valuation metrics to the proposed attack and validate the robustness and effectiveness of Truth-Shapley.