You ask how to connect to the VPN, and your agent starts filing an IT ticket. In agent development, that happens more often than you'd think. An internal docs agent got a new create_ticket tool. All the ticket examples passed and the average score went up. But comparing example by example showed that "How do I connect to the VPN?" now filed a ticket instead of searching the docs. The trace showed why: the tool description said "VPN and laptop issues," which matched every VPN-related question. Narrowing it to requests that need IT to act made every example pass. Looking only at the average would have hidden the regression. That loop is Eval-Driven Development. Like test-driven development, you define what good looks like as a spec and an eval first, confirm it fails on the current agent, then compare experiments example by example after each change. A held-out set guards against overfitting, and production traces keep growing the evals. What Is Eval-Driven Development? How to Improve AI Agents #AIAgents# #LLMEvals#

TL;DR: Classifier-style "decision models" can replace LLM judges at 16 to 357 times lower cost, but they get things wrong in the same places the LLMs do. Decision Models vs. LLMs as Rubric Judges: Cheaper, Faster, and Wrong in the Same Places Key points ・⚖️ Compares Jev and Perplexity's Decider against 3 flash-tier LLM judges on 7 benchmarks, 9 panels, 5,003 items ・💸 LLM judges cost 16 to 357x more and take 5 to 350x longer ・✅ Accuracy: significantly worse in at most 6 of 27 comparisons, and often stronger on binary panels ・📉 Weaker on ordinal scales, scoring lower than human raters ・🔁 When Jev is confidently wrong, 96.0% of LLM verdicts repeat the error (50.3% expected by chance) ・🪜 Cascades gain at most +2.7 points, and just +1.7 even with higher-tier fallbacks Swap them in to save cost; "defer to an LLM when unsure" is a weak accuracy safety net. #LLMasaJudge# #LLMEvaluation#

I’m at AIE NYC this week, and I’ll be giving a talk on evaluating subjective / open-ended tasks! Talk details: - Wednesday, October 14, 3:20pm - 3:38pm, Turing Room (Riverside Ballroom) - I’ll follow up with slides, links, full details, reading material, and more resources after the talk. Motivation. Evaluating LLMs is difficult, and much of this difficulty comes from noise. How do we know if a better eval result is due to a better model or just attributable to noise / variance? We need a methodology for building evaluations that yield confident / useful information despite the non-determinism of LLMs. Subjective tasks. The problem of noise in evaluations becomes even harder when working with subjective tasks. Recent LLM research focuses heavily on verifiable tasks, where correctness is judged with a deterministic check (e.g., string matches, symbolic verification, test cases, etc.). For subjective tasks, even defining correctness is an open question. We must clearly define what constitutes a correct solution and try to capture this with an LLM judge or rubric. Then, even our verifier is non-deterministic, thus making the evaluation process noisier. The subjectivity of the evaluation process makes it really easy to build poor evals that provide no useful information about model quality. Roadmap for Highly-Subjective Evals. To address these issues, my talk will outline a roadmap for properly evaluating subjective tasks. The roadmap consists of three primary components: 1. Human evaluation: the North Star for evaluation, must be calibrated and come with clear guidelines. 2. Offline (automatic) evaluation: an efficient (but imperfect) proxy for human evaluation that allows experiments to be scaled. 3. Production monitoring: key metrics that allow us to monitor quality in the wild and create “soft” model deployments via A/B testing. Multi-stage evaluation. These components create a tiered evaluation strategy for new models and ideas: Start with offline evaluation > Human evaluate promising candidates > A/B test in production > Deploy We have multiple stages of evaluation that each are used to detect different kinds of errors. By combining them together, we can evaluate model quality holistically, detect a much larger number of errors, and (hopefully) avoid surprises in production. Handling noise. Finally, there are a variety of techniques we can use to minimize the amount of noise in this evaluation process and increase confidence in our evaluation results: - Create clear quality guidelines for human evaluation. - Calibrate the human evaluation process to ensure agreement. - Calibrate LLM judges / rubrics to ensure agreement with humans. - Invest into proper data sampling for offline evaluation (diverse, realistic, difficult, and accurate). We should also leverage statistics at every step of the evaluation process. There are a core set of statistical concepts that can be used to quantify and reduce noise in LLM evals. We can use power analysis to estimate necessary sizes and setups for benchmarks / evaluation exercises, include confidence intervals with all of our evaluation results, re-sample model / judge completions multiple times to tighten confidence intervals, or even leverage pairwise model comparisons to more confidently detect differences in model evaluation results.

🧭 TL;DR: Instead of estimating an LLM's confidence from the current inference alone, this method calibrates it against how often similar past attempts actually succeeded. It beats 10-sample self-consistency in accuracy while costing far less. Title: Confidence Comes from Experience: Experiential Confidence Estimation from Reasoning to Agents URL: Points 🗂 Stores past task, reasoning trace, stated confidence, actual outcome, and a lesson learned in an experience bank 🔍 Recall stage retrieves k=50 similar episodes and computes their real hit rate statistically 💭 Reflect stage has the model restate confidence in words after seeing that track record 🏆 Matches or beats SC@10 on 23 of 24 model-dataset combinations by AUROC 🤖 Biggest gains on agent tasks where failures are silent — even beats a trained verifier on AppWorld 📉 2-20x lower calibration error across domains, at roughly 1/10th the compute cost A great example of the shift from "judge the current attempt" to "judge from accumulated track record." #LLMEval# #AIAgents#

Even with automated evals, only a person can tell why an agent failed. LangChain argues expert judgment is a strategic advantage 🧑‍⚖️ Why human insight is a strategic advantage for building better agents Automated evaluation can flag problems, but it can't explain the cause or the fix. The article lays out three principles for building human review into the evaluation system. 🎯 Principle 1: humans define what "good" looks like Reviewers supply context a trace doesn't contain, such as the standard of quality for the job and what the user was after. Scores with reasoning also expose gaps in the evaluators themselves. 🧵 Principle 2: context makes the judgment hold up Looking at one step misses contradictions with something said several turns earlier. For multi-turn agents, the whole conversation (a thread) is the unit of review. 🔁 Principle 3: turn judgment into evaluation data Labels become ground truth, used to align LLM judges with human scores, update online evaluators, and build golden datasets. In the case study, Fetch Rewards manually checks even product-card images and offer availability. Automation handles scale; the advantage comes from turning expert judgment into a reusable asset. #AIAgents# #LLMEvaluation#

How does an AI coach that can't touch medicine prove it's safe? Numan, a UK digital health company, built Nu, an AI coach that is a wellness product: no diagnosis, no prescriptions. It had to hold that line while still coaching well. The team first scored replies one by one on static datasets, but averages hid a few dangerous replies, and scripted conversations couldn't follow Nu's unexpected turns. Worse, some low scores came from a rubric demanding empathy and advice within two turns. The turning point was treating evaluation criteria, not prompts, as the scarce asset. They test the exact same LangGraph graph patients use, score by theme, let safety probes block regressions at release, and have a cheaper model play patients so a judge can score whole conversations. A provider migration that usually takes weeks took about a day. How Numan built an eval-first AI health coach with LangGraph and LangSmith #LangGraph# #LLMEvaluation#

💻 Even top-tier coding AIs see their scores collapse by more than half when the user isn't a programmer. A new benchmark puts that uncomfortable fact on the table. Title: SWE-Journey: Towards More Realistic Evaluation of Coding Assistants through Long-Horizon, Multi-Turn Interaction URL: ❓ What was the problem? Existing coding benchmarks tend to use short tasks with the full spec handed over up front, which doesn't match how real users gradually reveal requirements and react to what the assistant does. ❓ What does SWE-Journey do differently? It auto-synthesizes long-horizon tasks (averaging 10.2 changed files and 768 lines) and has AI play four personas drawn from real Claude conversation logs - non-coder, product manager, new developer, and software architect - each chatting with the assistant over dozens of turns. ❓ What did they find? Even the best model only reaches 52.22% Final F2P (final feature-implementation success). It hits 78.50% with the software-architect persona but drops to just 23.00% with non-coders, a 55.50-point gap. ❓ Why such a big gap? Weaker models rarely ask clarifying questions when requirements are vague, stay overconfident about the wrong part of the code, and sometimes report a fix as "done" with no real evidence. How well a model adapts its questioning and verification style to the user turns out to matter as much as raw coding skill. #AICoding# #LLMEvaluation#

Ever spend way too long hunting across scattered sources just to find "which benchmark actually fits this task"? Title: Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation URL: ❓ What is Benchmark Radar? 💡 It's a search engine that crawls 37 sources daily, arXiv, Hugging Face, GitHub and more, and pulls each benchmark's paper, dataset, repository, and score history into one place, with a CLI for offline queries. ❓ How much data are we talking about? 💡 1,283 source records from four catalogs, covering 12,916 numeric scores across 790 benchmarks. Daily collection alone logs 11,068 observations across 6,546 distinct artifacts. ❓ Why can't scores just be compared directly? 💡 Only 82 of the scored records use a verified 0-100 percentage scale, the rest use different or unverified scales, and roughly half lack a known release date, so naive comparisons would be misleading. ❓ So what's the fix? 💡 Instead of forcing comparability, it preserves source identity and citations so readers can inspect the evaluation conditions themselves. #LLMEval# #Benchmarks#

LLMs now beat humans on math accuracy. But does that mean they're actually building understanding on top of the right foundations? Title: Do LLMs Exhibit Coherent Knowledge Structures in Mathematical Reasoning? A Perspective from Knowledge Space Theory URL: ❓ Do LLMs actually build on prerequisite knowledge to get answers right? 💡 Accuracy favors LLMs (92.5% for the best model, Qwen3-80B, vs. 79.6% for humans), but "perfect prerequisite satisfaction" tells a different story: 72.7% for humans vs. only 48.16% for the best LLM. Lots of correct answers rest on shaky foundations. ❓ Does giving them prerequisite hints help? 💡 Surprisingly, prerequisite-grounded context barely outperformed unrelated examples. That points to surface-level pattern matching rather than genuine structured reasoning over prerequisites. ❓ Do strong models at least share a consistent knowledge structure with each other? 💡 Human learner groups overlap at 0.9+ in their knowledge structure, but LLM pairs only overlap 0.38-0.6 — and stronger models diverge even further from humans. ❓ So what's the takeaway? 💡 Accuracy alone hides how differently LLMs "understand" math. Knowledge Space Theory offers a lens that exposes the fragmented structure lurking behind impressive scores. #LLMEval# #MathReasoning#

A pipeline with 100% delivery rate, 100% schema validity, and zero retries or errors... yet re-running the exact same request flips the verdict. This paper reports that shocking negative result, fully preregistered with a complete audit trail. Title: Clean Engineering, Unstable Measurement: A Preregistered Reliability Failure of Black-Box LLM Observers on Shared Endpoints URL: 🔧 Highlight 1: Clean engineering does not mean reliable measurement Across 3,312 planned calls, response rate and schema validity both hit 100% with zero retries or errors, yet the repeat-ranking agreement (median Spearman) was only 0.400 — clearly failing the preregistered stability gate (threshold 0.90). 🎲 Highlight 2: Byte-identical inputs still drift Replaying the exact same request 24 hours later gave only a 0.780 exact-ranking agreement rate. Interestingly, comparisons across windows on the same day scored a nearly identical 0.805 median, revealing this isn't "next-day drift" but immediate platform-level nondeterminism. 📉 Highlight 3: More samples don't fix it Scaling observer calls from 8 to 500 (748,000 calls total) still left the stability gate pass rate at 0/500. The root cause: score gaps between candidates were 7-10 orders of magnitude smaller than the noise floor, making most tasks fundamentally unrankable. I think this is a great reminder that before using an LLM as a measurement instrument, you need to measure whether the instrument itself is stable. #LLMEval# #Reproducibility#