arXiv AI
Jul 21
Teach it to stop, not just to click
arXiv:2607. 17136v1 Announce Type: cross Abstract: Agentic computer-use RL is reported in single runs, and those numbers mislead.
Odds-Shift Slippage in One-vs-Rest Rankers: Diagnosing and Repairing Reweighting-Induced Top-K Errors
Related stories
arXiv AI
Oct 2
False Floors: LLM Safety Routing Evaluations Break Under Distribution Shift
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 Machine Learning
Aug 18
Whose Gold? Annotator-Pool Disagreement Is Large at the Item Level, and Hidden by Small Leaderboards
arXiv:2608. 15980v1 Announce Type: cross Abstract: Preference benchmarks are built by hiring annotators, and the identity of those annotators is treated as an implementation detail.
arXiv AI
Oct 7
Toward Alignment Scaling Laws: A Framework and First Preregistered Measurements
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 Machine Learning
Nov 2
Reliability-Weighted Direct Preference Optimization: Penalized Weights for Robust Preference Learning
arXiv:2610. 41588v1 Announce Type: cross Abstract: Direct Preference Optimization (DPO) treats every reported comparison as correct, yet preference data contain reversed, tied or otherwise noisy comparisons.
arXiv AI
Aug 10
Ratchet: How Reliable Must an LLM Judge Be to Retire a Skill?
arXiv:2605. 22148v3 Announce Type: replace Abstract: A large language model (LLM) agent that writes and edits its own skill library must also decide which skills to keep, from one noisy scalar per skill.
arXiv Machine Learning
Jul 8
Heckman-Corrected Epistemic Uncertainty: Selection on Unobservables Defeats Importance Weighting
arXiv:2607. 05806v1 Announce Type: new Abstract: Training data for machine learning is routinely collected by a selection process the model never sees: loans are observed only when granted, outcomes only when a test was ordered.
arXiv Statistics ML
Oct 27
Identities, Not Pairs: Distribution-Free False-Accept Control for Embedding-Based Verification
arXiv:2610.35277v1 Announce Type: new Abstract: Embedding-based verification often calibrates a false-accept threshold from all impostor pairs among $m$ identities, although these pairs share identit...
arXiv Machine Learning
Jul 21
The Label Complexity of Class-Conditional Coverage under Distribution Shift
arXiv:2607. 18088v1 Announce Type: new Abstract: Standard evaluation of many recognition systems contains distribution shift by construction, since benchmarks place disjoint conditions in the training and test splits.
arXiv AI
Oct 27
Audit Where Hacks Pay: Budgeted Strong-Verifier Audits Against Reward Hacking in GRPO
arXiv:2610.35716v1 Announce Type: cross Abstract: Reinforcement learning with verifiable rewards trains on cheap checks, such as unit tests, that a policy can learn to game, while a stronger check is...
arXiv Machine Learning
Sep 25
Optimal Recovery Meets Bayesian Learning: Where Worst-Case Bounds Pay Off
The paper shows that Worst‑Case Optimal Recovery (OR) and Bayesian learning solve the same Gaussian‑quadratic‑Hilbert problems, linking the radius of information to a nugget‑optimized Gaussian process posterior variance. It evaluates three Bayesian systems, demonstrating that OR can outperform Bayesian methods in certain calibration and reproducibility metrics, yet split‑conformal and other approaches can beat OR in interval scoring, especially under covariate shift. The authors propose matching the guarantee tool to the data regime and auditing that regime first.
arXiv AI
Aug 25
Crossing the Margin Cliff: Toward Relearn-Robust LLM Unlearning via Margin Calibration
arXiv:2607.27836v2 Announce Type: replace Abstract: Large language model unlearning is consistently fragile under relearn attacks. On TOFU, fine-tuning on twenty forget examples substantially recover...