Concept erasure, or concept unlearning, in text-to-image diffusion models is usually audited with prompt attacks. A prompt attack tests whether one conditioning string still retrieves the erased concept. It does not measure the probability that the frozen generator still assigns to that concept. We introduce least-action auditing of this residual probability, and evaluate it on Stable Diffusion concept-erasure checkpoints. The least-action barrier is the minimum control cost of driving the frozen reverse process into a detector-defined concept event. This cost equals the path-space KL divergence of conditioning the model on the event, and therefore the event's negative log-probability. The estimate $ΔS_{\mathrm{rob}}$ uses a CLIP$\wedge$DINOv2 consensus terminal, with no retraining and no adversarial prompt. Across $18$ erased concept-mechanism pairs spanning object classes and NSFW nudity, $ΔS_{\mathrm{rob}}$ predicts deployment leakage counted on $1{,}500$ fresh samples per model by detectors excluded from the audit (pooled Spearman $ρ{=}{-}0.76$, $p{=}5{\times}10^{-4}$). Within every audited concept, the eraser with the higher barrier leaks less ($18/18$). On NSFW nudity the barrier orders the standard erasers exactly: UCE $ΔS_{\mathrm{rob}}{=}3.9$ (leak $16.4\%$) $<$ ESD $75.5$ ($9.8\%$) $<$ RECE $224.5$ ($1.4\%$), while Safe Latent Diffusion (SLD) stays near zero and leaks $85.5\%$ under explicit prompts. The same trajectories repair deployed checkpoints. An audit-guided closed-form patch reduces UCE nudity leakage from $16.2\%$ to $0/1{,}500$ and RECE church leakage by $3.3\times$, where direction-count-matched random edits fail.
Motion that is plausible on one body can penetrate another after a change in shape, and motion models can produce self-penetration on their native characters. We present a training-free post-process that reduces these artifacts through pose correction followed by bounded skin compression. The module attaches to any motion model whose native character and output character have compatible character packs, without access to model internals. Each pack is built offline from the rig and skin without training. A vertex-level pose layer solves a local quadratic program in plain C on one CPU core, and a stateless contact deformer acts on the host's final skinned mesh. A part-dependent soft-tissue budget allocates correction between pose and skin through a static convention. Source penetration contributes no extra allowance; optional cloth thickness and compliance modify the pose layer's collision surface. Operation is causal or uses a short buffered lookahead, with frame-exact pass-through at the native interface for clean, inactive frames in the zero state. Across 320 unseen clips and 3 body shapes, the fraction of frames with hand-torso depth above 20 mm falls from 15.53% for copied motion to 0.00%, with final maximum depth 9.3 mm over every frame. Mean pose and skin costs are 1.11 and 1.18 ms, respectively, and their sum of 2.29 ms is a sub-frame processing cost. The median joint-velocity high-frequency ratio relative to copied target motion is 1.008.
LLMs can predict human survey responses with nontrivial accuracy. What this means for the future of social surveys remains hotly contested. Optimists anticipate a coming era of synthetic respondents; skeptics see a fragile and overhyped substitute. This article advocates a position: AI predictions cannot replace surveys, but they can augment them. Using prediction-powered inference (PPI), a statistical framework rooted in survey methodology and known in sociology as "mixed subjects design", the study tests whether AI predictions can improve the statistical efficiency of social surveys when combined with census-scale auxiliary information. Applying the framework to 83 social outcomes in the General Social Survey (GSS) and National Health Interview Survey (NHIS), with American Community Survey (ACS) being the auxiliary file, we find that supervised predictors yield effective-sample-size gains of about 7-12 percent when estimating the mean, while LLM gains are smaller but positive.
Large Reasoning Models (LRMs) embedded in agentic frameworks have transformed information retrieval from static, long-context question answering into open-ended exploration. Yet real-world use requires models to discover and synthesize “long-tail” facts from dispersed sources, a capability that remains under-evaluated. We introduce PolitNuggets, a multilingual benchmark for agentic information synthesis via constructing political biographies for 400 global elites, covering over 10000 political facts. We standardize evaluation with an optimized Supervisor–Searcher multi-agent system and propose FactNet, an evidence-conditional protocol that scores discovery, fine-grained accuracy, and efficiency. Across models and settings, we find that current systems often struggle with fine-grained details, and vary substantially in efficiency. Finally, using benchmark diagnostics, we relate agent performance to underlying model capabilities, highlighting the importance of short-context extraction, multilingual robustness, and reliable tool use.
cs.CL · cs.AI · cs.MA
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表