Part of our AI Eval Watch series · Nov 4–10, 2026
This week's evaluation research kept returning to one measurement question: when a score looks good, what has it actually measured? Below are three preprints and one independent report. The preprints are not peer reviewed, and each item describes what its authors report, not something we have independently verified.
Hosted LLM judges tend to track human ratings well, but they cost money per call and can be hard to use where data has to stay in-house. Small open-weight judges are cheap to run but usually agree with people less. The authors try to close that gap with soft prompt tuning: they learn prompt embeddings from the human labels in FairBotBench, in some versions combined with ordinary written instructions, and then test the tuned judges on three aspects of chatbot replies: bias, toxicity and personalization. The hybrid version, written instructions plus learned embeddings, raised agreement with the human annotations over untuned baselines in every comparison. For Qwen 3.5 9B, the tuned judges beat both hosted reference judges on all three aspects.
Why it matters. This is the useful kind of validity evidence: the judge is checked against human ratings, not against other judges. Two limits are worth keeping in mind. The abstract does not give the size of the gains or which agreement statistic was used, so it is not yet possible to tell whether the improvement is large. And a judge tuned toward one team's labels learns that team's standard, blind spots included, so its agreement with your annotators can only be as good as your annotators' agreement with each other. Measure that first. Our guide on what makes an AI eval defensible to auditors covers checking a scoring method against a human baseline.
The authors study vision-language models facing compositional harm, where an image and a prompt are each harmless alone but harmful together. Safety benchmarks in this area are scored by an LLM judge that reads the model's replies. The authors first show that a small supervised fine-tune reaches about the same judge score as a multi-stage reinforcement-learning pipeline. They then show that the safe behavior behind such scores can be removed by a targeted change to the model's internals: on two open models, safety scores fell by up to 65 points while more than 98% of outputs stayed coherent.
Why it matters. A judge that reads outputs can only report how a model behaved on the test prompts. Two systems can earn the same score and still differ sharply in how well that behavior would hold up under pressure. In measurement terms, the score can be reliable and still not be valid for the claim people tend to draw from it, which is that the model is robustly safe. Validity belongs to an interpretation of a score, not to the number itself, so before a score supports a decision, ask what other evidence would have to agree with it. Our guide on evaluating an LLM beyond average accuracy covers what a single headline number leaves out.
This one comes from chemistry. One LLM pulls reaction records out of published papers, citing the passages it used. A second, independent LLM judge then checks each record against the full paper, flags fields the paper does not support and proposes cited corrections. To test the judge, the authors compared it with manual review on a hand-curated reference set of 295 experiments. It matched the manual review on 85.7% of the records and was better than a simple string-matching check at spotting mistakes that spanned related experiments. Run on 3,942 records from 561 papers, the judge accepted 67.5% as written. Most of its fixes were to numbers, where the extractor had mixed up values from neighboring experiments. The authors release the dataset together with the judge's audit logs.
Why it matters. Here a validity check exists: the judge is compared with human-curated reference data. Its limits are worth stating. The 85.7% is raw agreement. The abstract gives no chance-corrected statistic and no breakdown of where the disagreements fall. With 295 records, a rough 95% margin of error on that figure is about 4 percentage points either way, assuming the records are independent. The reference set also covers one narrow sub-area, so the figure may not carry over to all 3,942 records. Releasing the audit logs is the right call, because it lets others check the judge's decisions. Our guide on how many test items an eval needs shows how sample size sets that margin.
METR published an independent assessment of AI research and development acceleration at Anthropic. One part of it tracks Anthropic's internal capability index, which is built from benchmark scores. METR estimates that the index has grown about 1.9 times faster since October 2025 than before, with a 95% interval of 1.5 to 2.8. The interval comes from a bootstrap that resamples benchmark scores within each model, combined with a trend line that has a single breakpoint. METR notes that the interval covers uncertainty in the benchmark scores only, not uncertainty about which models were included. Under other reasonable analysis choices the estimate runs from 1.1 to 2.9 times, driven mostly by the choice of link function, and adding other labs' frontier models lowers it to 1.4. METR also raises saturation: the newest internal model scored above 50% on all 15 benchmarks it was tested on, which METR says worries it about future measurement, although capping benchmarks at 85% left the estimate at 1.9. Anthropic redacted much of the evidence METR considered most informative.
Why it matters. This is careful reporting, and it shows a general point. The spread across reasonable analysis choices (1.1 to 2.9) is wider than the confidence interval (1.5 to 2.8), so a single interval understates the total uncertainty. When you report an improvement, show how much the result moves under other defensible choices as well as the interval. Saturation is a reliability problem: once most systems score near the top of a benchmark, it stops separating them, and measurement noise makes up more of any remaining gap. Our guide on telling whether one LLM is really better than another walks through checking whether a gap is bigger than the noise.
If you are building evaluation metrics for an AI product and need them to hold up under real scrutiny, this is squarely what we do.
One email whenever we publish something new. No spam, unsubscribe anytime.
{{ subscribeErrorMsg }}