Searcharxiv⌕ Search

arXiv subjects

Ook Lee

Publications and source records attributed to Ook Lee.

3 recordsLinked to original sources

Room to Reconsider: Reply Length Decides Whether Pushback Makes Language Models Sycophantic or Self-Correcting

Language models often abandon a correct answer when a user pushes back without giving a reason, and this behavior is usually studied as sycophancy to be suppressed. The same pushback, however, also leads models to revisit and sometimes repair a wrong answer. We measure both effects on multiple-choice questions from MMLU and ARC-Challenge with four open models from two families (Qwen3 1.7B, 4B and 8B, and OLMo-2 7B), reading the final answer exactly from option logits after a free-form reply. The reply budget changes the verdict. When Qwen3-4B may answer the challenge in at most 96 tokens, pushback leaves accuracy unchanged or slightly lower; with 1,024 tokens, it corrects 43% of its wrong answers and pushback raises accuracy by 5.7 points. Whether pushback helps overall follows a simple break-even rule in the cave and correction rates. Inside the model, caving can be partly predicted before the challenge: hidden states add about 0.04-0.06 AUROC over turn-1 confidence in three of our four probing runs. A probe trained only on self-correction cases predicts caving beyond confidence when replies are truncated, but not when the model can finish its reasoning. Activation steering never reduced caving selectively: where it worked, it suppressed self-correction as well. We conclude that sycophancy evaluations should allow full replies and report self-correction alongside caving.

cs.CL↗

When Is Disclosing Private Conversations with AI to the State Justified? Proactive Referral Policies of Major U.S. AI Companies through the Lens of Contextual Integrity

AI companies engage in "proactive referral," notifying law enforcement of user conversations they judge dangerous without a government request. This poses a dilemma, since referral may expose people who merely confided dangerous thoughts to an AI to investigation, whereas not referring may leave actual violence unprevented. This study asks under what conditions proactive referral can be justified and to what extent companies' public policies disclose those conditions. Drawing on contextual integrity theory and the norms governing exceptional information sharing in confidential counseling, online posts, and private communications, the study develops eight justification criteria classified as required, presumptive, or recommended. Three evaluators analyzed 27 public documents from six major U.S. AI companies using the criteria applicable to corporate policies. No company clarified whether broad sharing provisions or narrower safety notices took precedence, and assessments of referral scope depended on which provisions were given precedence. No company disclosed the specific factual thresholds and exception conditions governing referral decisions, or whether users would be notified afterward. The study turns contextual integrity into a stage-by-stage framework for proactive referral, shows that policy documents must be read in light of their relationships, and proposes policy elements for keeping proactive referral a narrow exception disclosed in advance.

cs.CY↗

Bridging KAN and MLP: MJKAN, a Hybrid Architecture with Both Efficiency and Expressiveness

Kolmogorov-Arnold Networks (KANs) have garnered attention for replacing fixed activation functions with learnable univariate functions, but they exhibit practical limitations, including high computational costs and performance deficits in general classification tasks. In this paper, we propose the Modulation Joint KAN (MJKAN), a novel neural network layer designed to overcome these challenges. MJKAN integrates a FiLM (Feature-wise Linear Modulation)-like mechanism with Radial Basis Function (RBF) activations, creating a hybrid architecture that combines the non-linear expressive power of KANs with the efficiency of Multilayer Perceptrons (MLPs). We empirically validated MJKAN's performance across a diverse set of benchmarks, including function regression, image classification (MNIST, CIFAR-10/100), and natural language processing (AG News, SMS Spam). The results demonstrate that MJKAN achieves superior approximation capabilities in function regression tasks, significantly outperforming MLPs, with performance improving as the number of basis functions increases. Conversely, in image and text classification, its performance was competitive with MLPs but revealed a critical dependency on the number of basis functions. We found that a smaller basis size was crucial for better generalization, highlighting that the model's capacity must be carefully tuned to the complexity of the data to prevent overfitting. In conclusion, MJKAN offers a flexible architecture that inherits the theoretical advantages of KANs while improving computational efficiency and practical viability.

cs.LG↗