arXiv:2607. 02587v2 Announce Type: replace-cross Abstract: Trust-benchmark scores reported on a chat-LLM release line are often carried across several checkpoints of the same line, as if the underlying model had not shifted between releases.
arXiv:2607. 02587v1 Announce Type: cross Abstract: Model cards quote trust-benchmark scores without recording when they were measured, and the same number is routinely carried across successive checkpoints of one release line as if the model behind it had not shifted.
ProxyGuard is a new method for assessing the reliability of randomized data release mechanisms that use shared target sets. It offers two modes: named-release mode, which corrects for multiplicity and certifies specific releases, and direct shared-target mode, which evaluates independent mechanism draws on a common target, providing finite-sample reliability guarantees without needing independent target batches. In a registered study, direct mode increased power from 5.6% to 64.2% at a 0.95 reliability level, while named mode performed better under high-signal evidence.
arXiv:2609.13714v1 Announce Type: new Abstract: An updated model can improve an aggregate metric while degrading a slice that matters to a downstream user. We study checkpoint selection subject to no...
The paper introduces LSREP, a Longitudinal State‑Replay Evaluation Protocol designed to assess how conversational memory evolves over time, incorporating ordered replay, lifecycle schedules, repeated probes, evolving reference answers, and mechanism‑fidelity checks. It applies LSREP to ICE v2, a local‑first memory middleware, and reports that on three ordinary‑density datasets ICE v2 achieves near‑zero mean quality difference from vector‑RAG while using fewer fragments but slightly more prompt tokens, yet fails catastrophically on a dense dataset. In a public diagnostic, ICE v2 underperforms pure vector‑RAG on LongMemEval, revealing significant multi‑session and temporal failures and a quality‑cost trade‑off rather than superior efficiency.
arXiv:2608. 02685v1 Announce Type: cross Abstract: Coding-agent benchmarks increasingly cover long-horizon, end-to-end, and interactive development, but typically retain one requested outcome or a fixed change sequence.
The study evaluates whether prompt‑token counts can reliably identify the lineage of large language models served via APIs. Using a frozen‑threshold approach on 24 labeled endpoint pairs, the authors find that token‑count consistency perfectly separates development pairs but only half of the holdout pairs meet the strict repeatability criteria, yielding moderate accuracy and perfect specificity. The results confirm token‑count consistency as a fingerprint of shared tokenization stacks but reject it as a standalone test for model‑family attribution.
The paper argues that evaluating continual knowledge‑updating methods solely at a final checkpoint and a single adapter rank can be misleading. By fixing a periodic hierarchy and comparing it to cumulative replay on a 24‑month Wikidata stream, the authors show that the apparent best method changes depending on the evaluation month, replay LoRA rank, and query formulation. They recommend reporting performance trajectories and capacity sweeps, and only declaring a robust winner when the ranking remains stable across the evaluation region.
The paper investigates how persistent memory in AI agents can lead to over‑trust in stale facts, creating a "Memory Trust Gap" that worsens as model capability increases. Using a benchmark with Benefit and Safety suites across Qwen3 models of varying sizes, the authors show that larger models are more prone to harmful over‑trust, especially when metadata is absent or misleading. They also demonstrate that mitigation strategies such as exposing metadata or pre‑resolving conflicts improve accuracy, but the effectiveness depends on model size and dataset.
The paper investigates the reliability of ranking tables produced by small-sample evaluations of large language models (LLMs). Using LLM‑inferred prompt structure across eight model variants, the authors find that prompt‑structure recovery is highly unstable, with only the bottom of the ranking consistently reproducible. They demonstrate that standard evaluation practices can misrepresent model performance and propose reporting practices to improve transparency.
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.
AutoTuneBench introduces a trustworthy measurement protocol for evaluating how large language model agents auto‑tune GPU kernels and serving engines. The benchmark addresses four failure modes—strawman baselines, machine‑dependent timing, saturated tasks, and infrastructure defects—by enforcing code‑frozen protocols, database validation, anti‑cheat checks, pre‑registered comparisons, and external result anchoring. Using this protocol, the authors demonstrate that previously reported speedups are inflated, revealing more modest improvements across different engines and machines.
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.