arXiv:2603.12564v5 Announce Type: replace-cross Abstract: Tool-augmented LLM agents increasingly operate as multi-turn advisors in high-stakes domains, yet their evaluation relies on ranking metrics that measure what is recommended but not whether it is safe for the user. We present a paired-trajectory protocol that replays real financial dialogues under clean and contaminated tool-output conditions across eight LLMs (7B to frontier), decomposing divergence into information-channel and memory-channel mechanisms. We observe evaluation blindness: recommendation quality is preserved under contamination (UPR~1.0) while risk-inappropriate products appear in 65-93% of turns, invisible to standard NDCG. Violations are information-channel-driven, emerge at turn 1, and persist without self-correction over 23-step trajectories. Even non-extreme perturbations (within-band corruption, narrative-only attacks) evade threshold monitors while producing significant drift. Susceptibility scales with instruction-following fidelity across all eight models. Sparse autoencoder probing reveals models internally distinguish adversarial perturbations but fail to propagate this signal to output; causal interventions (activation patching, feature clamping, direct steering) confirm this representation-to-action gap is structural and resists linear repair. A safety-penalized NDCG variant (sNDCG) reduces preservation ratios to 0.51-0.74. These results motivate trajectory-level safety monitoring for deployed multi-turn agents.

arXiv:2603.23043v1 Announce Type: cross Abstract: The accelerating pace of climate change introduces profound non-stationarities that challenge the ability of Machine Learning based climate emulators to generalize beyond their training distributions. While these emulators offer computationally efficient alternatives to traditional Earth System Models, their reliability remains a potential bottleneck under "no-analog" future climate states, which we define here as regimes where external forcing drives the system into conditions outside the empirical range of the historical training data. A fundamental challenge in evaluating this reliability is data contamination; because many models are trained on simulations that already encompass future scenarios, true out-of-distribution (OOD) performance is often masked. To address this, we benchmark the OOD robustness of three state-of-the-art architectures: U-Net, ConvLSTM, and the ClimaX foundation model specifically restricted to a historical-only training regime (1850-2014). We evaluate these models using two complementary strategies: (i) temporal extrapolation to the recent climate (2015-2023) and (ii) cross-scenario forcing shifts across divergent emission pathways. Our analysis within this experimental setup reveals an accuracy vs. stability trade-off: while the ClimaX foundation model achieves the lowest absolute error, it exhibits higher relative performance changes under distribution shifts, with precipitation errors increasing by up to 8.44% under extreme forcing scenarios. These findings suggest that when restricted to historical training dynamics, even high-capacity foundation models are sensitive to external forcing trajectories. Our results underscore the necessity of scenario-aware training and rigorous OOD evaluation protocols to ensure the robustness of climate emulators under a changing climate.

arXiv:2603.12564v1 Announce Type: cross Abstract: Tool-augmented LLM agents increasingly serve as multi-turn advisors in high-stakes domains, yet their evaluation relies on ranking-quality metrics that measure what is recommended but not whether it is safe for the user. We introduce a paired-trajectory protocol that replays real financial dialogues under clean and contaminated tool-output conditions across seven LLMs (7B to frontier) and decomposes divergence into information-channel and memory-channel mechanisms. Across the seven models tested, we consistently observe the evaluation-blindness pattern: recommendation quality is largely preserved under contamination (utility preservation ratio approximately 1.0) while risk-inappropriate products appear in 65-93% of turns, a systematic safety failure poorly reflected by standard NDCG. Safety violations are predominantly information-channel-driven, emerge at the first contaminated turn, and persist without self-correction over 23-step trajectories; no agent across 1,563 contaminated turns explicitly questions tool-data reliability. Even narrative-only corruption (biased headlines, no numerical manipulation) induces significant drift while completely evading consistency monitors. A safety-penalized NDCG variant (sNDCG) reduces preservation ratios to 0.51-0.74, indicating that much of the evaluation gap becomes visible once safety is explicitly measured. These results motivate considering trajectory-level safety monitoring, beyond single-turn quality, for deployed multi-turn agents in high-stakes settings.