Reliability

TL;DR for operators An LLM judge can agree closely with human reviewers, produce consistent scores, and still provide insufficient evidence for a product decision. Agreement measures correspondence between evaluation methods; it does not establish that either method adequately captures the quality or risk the decision requires. LLJ Cards proposes five documentation and evaluation components, containing 45 guiding questions, to address this gap. The framework makes evaluation context, measurement criteria, prototyping, pipeline specifications, and validity evidence explicit. ...

TL;DR for operators Enterprise AI approval should attach to the complete workflow configuration, not to a model name or benchmark score. EnterpriseVal formalizes that idea by freezing the model, prompts, retrieval or memory, tools, guardrails, logging, and human-review design before evaluation. The framework then tests six families of evidence—fidelity, utility, efficiency, reliability, assurance, and oversight—and uses precommitted thresholds with uncertainty to classify a workflow as Reject, Conditional, or Scale. Only candidates that survive those risk gates should be ranked, and the preferred candidate can change when the business changes the relative importance of its metrics. ...

TL;DR for operators A persistent-memory agent can retrieve two harmful records that substitute for one another. Remove either record alone, and the remaining one can preserve the same wrong answer, making both look harmless under a one-at-a-time deletion test. MeClear1 addresses that failure mode by measuring each retrieved memory across multiple combinations of the surrounding context rather than only against the full retrieved set. It uses sampled Shapley contributions to identify negatively contributing memories, ranks them, and then tests nested clearance candidates. A nonempty clearance is accepted only when it produces positive empirical task gain and passes a fixed recovery check. ...

TL;DR for operators A deterministic verifier can tell you whether an artifact satisfies a formal specification. It cannot tell you whether that specification faithfully represents what the user originally asked for. Magenta makes this distinction explicit. The system first converts a natural-language mathematics problem and proposed answer into a Lean statement, then uses a learned judge to check whether that statement preserves the intended problem before attempting a machine-checked proof. When proof verification fails, another judge classifies the failure as mathematical or formal-implementation related and sends the next correction to the corresponding component.1 ...

TL;DR for operators A simulator can clear an agent while the physical deployment simultaneously fails one operational claim and satisfies another. In the paper’s 240-trial calibration block, every source-cleared execution was a false clearance for immediate completion, while only 20 of 240 failed on reported state, 42 of 240 failed on observable physical effect, and none failed on the eventual settled postcondition. ...

TL;DR for operators A tool-using agent can remember the right customer, constraint, or prior result and still make the wrong call. From State to Action: OODA-Tool for Reliable Multi-Turn Tool Use1 treats that gap as an architectural problem rather than only a prompting problem. Its strongest configuration separates four jobs: reconstruct the active task state, decide whether execution is actually warranted, choose a permitted action structure, and only then bind concrete arguments. On ToolDial, Specialized OODA beats Direct-LoRA at every tested Qwen3 scale, by 4.48 to 6.99 percentage points in Task Success. The gains are largest where state must survive long histories, missing information, changed values, constraints, or sequential dependencies. ...

TL;DR for operators When one model answer can trigger a consequential downstream step, spending more inference compute before trusting that answer can help—but the form of that spending matters. In the reported experiment, generating several reasoning attempts and aggregating the answer that repeatedly emerged increased verified acceptance from 56.2% to 64.9%. Asking the same model to critique and revise itself produced a smaller increase, from 47.2% to 50.6%. Adding a second model did not improve the reported acceptance rate: 47.4% of outputs survived cross-model verification versus a 48.7% single-model baseline. ...

TL;DR for operators A model that is highly reliable at applying one explicit rule transition is not necessarily reliable at executing an entire procedure built from those transitions. In Reasoning Capabilities of Large Language Models. Lessons Learned from General Game Playing1, the strongest evaluated model, Gemini 2.5 Pro, achieves 95.6% exact success on one-step next-state generation. At five dependent state transitions, exact success falls to 73.4%. When the model must also choose actions during those five steps, it falls again to 65.3%. ...

TL;DR for operators When one model output becomes input to the next stage, a final accuracy score tells you too little about where reliability is being lost. A workflow may fail because a required fact was never available, because a later composition step is intrinsically harder, or because an earlier mistake was allowed to propagate. Those failure modes call for different controls. ...

TL;DR for operators A model saying “8/10 confident” does not mean its underlying uncertainty is approximately 20%. Across the evaluated settings, the average instance-level correlation between reported confidence and logits-based confidence is only 0.135. The more useful operating rule is narrower. First test whether reported scores vary enough to distinguish cases. Then measure whether those scores rank examples meaningfully on held-out data. Separately test whether their numerical scale agrees with the comparison signal and whether they are calibrated against correctness. Do not substitute one test for another. ...