arXiv:2607. 01854v1 Announce Type: cross Abstract: Can a platform tell, before deployment, whether an open-weight checkpoint has had its refusal mechanism stripped?
arXiv:2607. 09800v2 Announce Type: replace Abstract: Master weights and stochastic rounding bypass invisible stored-weight updates but do not locate lost direct-storage proposals or parameters worth protecting.
The paper evaluates hard‑gate candidacy for validators in a deployed generative‑agent system by measuring how well each validator’s firing separates successful from failed builds. Across 13 validators and thousands of builds, only a few checks show statistically significant separation, while many fail to distinguish or never fire. The study highlights that skipped checks are recorded as passes, limiting detectable failure rates and underscoring the need for clearer evaluation records.
arXiv:2609.06934v1 Announce Type: cross Abstract: Post-hoc safety training (RLHF, DPO) is the dominant way to align large language models, yet jailbreaks (Zou et al., 2023b), fine-tuning attacks (Qi...
arXiv:2608. 20290v1 Announce Type: new Abstract: Whether a language model has improved itself is increasingly judged not by mean accuracy but by which individual problems it gains and loses.
arXiv:2608. 05863v1 Announce Type: new Abstract: Modern models no longer keep a plain KV cache: latent caches, learned sparse selectors and recurrent states each carry the model's memory in a different form, and each fails differently under compression.
RAGScope is a leakage‑controlled, cost‑aware protocol that evaluates evidence‑gating mechanisms for retrieval‑augmented generation (RAG) systems using only the task input, retrieved context, and answer text. The enhanced gate, RAGScope‑E, achieves an AUROC of 0.798 and an average precision of 0.660 on three RAGTruth tasks, outperforming ROUGE‑L by 0.034 in pooled AP and delivering 0.748 precision within a top‑10% review budget. It operates quickly (6.22 ms per example on CPU) and demonstrates that cheap evidence gates can effectively triage RAG outputs, though calibration must be validated and adapted for each target domain.
The paper reports a preregistered audit of language‑model judges used as measurement instruments, revealing that the assumption that a model’s responses remain stable over time is invalid. Across nearly 53,000 audited requests, repeat rankings and byte‑identical replays fell far below required reliability thresholds, with three identified mechanisms—label‑to‑meaning bias, candidate gaps below the noise floor, and input permutation noise—explaining the discrepancy. The study proposes a three‑level snapshot‑identity framework, eight design rules, and a reporting checklist to prevent such reliability failures in future evaluations.
arXiv:2606. 10154v1 Announce Type: new Abstract: Quantized checkpoints are often screened first with quality metrics and only later, if at all, with direct safety tests.
The paper investigates the reliability of language‑model judges used as measurement instruments on shared endpoints. Through two preregistered audits of 52,988 requests, the authors found that repeat rankings and byte‑identical replays fell far short of required thresholds, revealing significant instability. They identify three mechanisms—label‑to‑meaning bias, candidate gaps below the noise floor, and input permutation noise—that explain the gap, and propose a snapshot‑identity ladder, design rules, and a reporting checklist to mitigate such failures.
arXiv:2608. 10145v1 Announce Type: new Abstract: LeWorldModel trains a latent world model with a prediction loss and a single anti-collapse regulariser, and reports approximately 87% of goals reached on TwoRoom, its simplest diagnostic environment.
The paper introduces a claim‑safe protocol for evaluating closed‑loop AI systems, consisting of three actions: Refuse, Decompose, and Refresh. It demonstrates the protocol in a simulator with 24 policy components and 1,440 held‑out cases, showing that abstention and stable false admission rates are low while providing detailed statistical diagnostics. The approach emphasizes that evaluation results should be tied to observable support and statistical calibration rather than a single PASS/FAIL label.
arXiv:2608. 15286v1 Announce Type: cross Abstract: We introduce AgentRelBench, an environment-agnostic reliability instrument that computes ground-truth, severity-priced damage from database state diffs across repeated runs, with no LLM in the measurement path, demonstrated on EnterpriseOps-Gym.