The paper proposes a framework for studying how alignment difficulty scales with model size, treating alignment as a family of measurable scaling relations for different risk categories. It introduces operational definitions of burden, distinguishes observed, audited, and true alignment, and proves that the largest exponent among corrected risks determines long‑run behavior. The authors preregistered and applied the protocol to public data, finding scaling exponents for attack success, truthfulness, dispositions, sycophancy, and a backdoor, and released interactive browser games to illustrate the laws.
arXiv:2606. 30627v1 Announce Type: cross Abstract: Conservative offline training is widely advocated as a safe foundation for subsequent online adaptation: if a policy stays close to well-supported behaviour, the argument goes, it is less likely to exploit imperfections in a learned reward model.
The paper examines safety routers—systems that route user requests to different language models—and finds that their performance degrades significantly when evaluated under distribution shift. In standard benchmarks, routers appear effective because the best single model is chosen from the same evaluation data, but when the data distribution changes, the routing advantage diminishes or disappears. The study quantifies this bias across multiple safety corpora, showing that routers offer little benefit under realistic shift conditions and that recognition‑based defenses can be undermined by attackers who know the model being used.
arXiv:2607. 09706v1 Announce Type: new Abstract: Language models turn a worded situation into a numeric plan, and the dominant pipelines (NL4Opt, OptiMUS, ORLM, OR-LLM-Agent) commit to a single objective and point-valued coefficients, then solve once.
arXiv:2607. 18292v1 Announce Type: cross Abstract: As language models scale, answers start truer but degrade faster: scaling buys capability but erodes reliability.
The paper introduces a new closed‑form scaling law that extends Chinchilla’s original formula to handle data‑constrained regimes. It decomposes loss into undercapacity, undertraining, and overfitting components, saturating between an irreducible loss and an uninformed baseline. The authors validate the model on diverse architectures and domains, achieving state‑of‑the‑art RMSE across multiple LLM scaling‑law grids and enabling cost‑aware training allocations.
arXiv:2606. 08919v1 Announce Type: new Abstract: As LLM agents begin to take real, irreversible actions (shell commands, file edits, deploys), the standard safety pattern is a human-in-the-loop approval gate: risky actions pause and wait for a person.
arXiv:2601. 22313v2 Announce Type: replace Abstract: Large Language Models (LLMs) are rarely static and are frequently updated in practice.
arXiv:2606. 29623v1 Announce Type: new Abstract: Rare events govern the safety profile of modern AI systems, yet their probabilities are extremely difficult to estimate: direct Monte Carlo requires prohibitive sample budgets.
The paper introduces DISCERN, a two-tier protocol for certifying that updates to production models do not increase risk. It first uses unlabeled data to detect benign updates based on disagreement rates, then selectively labels only disagreements through an anytime-valid confidence sequence. The method achieves finite-sample validity with label-complexity bounds of order ρ²/ε², demonstrating significant label savings and strong empirical performance across 14,000+ audit streams.
arXiv:2607. 18292v2 Announce Type: replace Abstract: As language models scale, answers start truer but degrade faster: scaling buys capability but erodes reliability.
The paper introduces Alignment Forecasting, a method for predicting whether fine‑tuning a language model on a given dataset will increase specific alignment failures such as deception or sycophancy. It presents ALIGNMENTFORECASTBENCH, a benchmark of over 5,000 forecasting questions across many models, datasets, and failure modes, and shows that a simple forecasting scaffold using an LLM’s assessment of dataset bias can outperform baseline forecasters. The authors demonstrate that filtering out high‑risk training examples identified by the forecaster can improve alignment in multiple‑choice evaluations, though benefits in open‑ended conversations remain uncertain.
arXiv:2605.01699v4 Announce Type: replace Abstract: Recent attacks show that behavioural unlearning of large language models leaves internal traces recoverable by adversarial probes. We characterise...