AI reliability: repairing failures without rewriting the task

Reliability depends on what gets repaired

This week's selected papers connect through a question: when an AI system encounters failure or ambiguity, what should change—the model, the workflow, or the interpretation of success? Safety finetuning and agent failover deliberately alter how systems respond; a study of grading rules shows why silently repairing instructions can instead undermine fidelity. These are related problems, not evidence of agreement on one solution. Addition by Subtraction, MA-Mesh and One Scale, Four Texts operate at different levels.

Hardening the model versus reorganizing the work

Addition by Subtraction temporarily deactivates identified safety-critical neurons during preference-based finetuning, encouraging safety behavior to persist through other model components. Across four models from three vendors, the authors report improved resistance to pruning, refusal-direction ablation and black-box jailbreaks. But resistance to these interventions does not directly measure the absolute density of safety representations: neuron identification is imperfect, and removing multifunctional neurons risks capability damage. “More distributed” is therefore an operational interpretation of reduced fragility, not a complete map of safety inside the model. Addition by Subtraction

MA-Mesh instead adds a runtime recovery layer: agents assess peers against role-specific requirements, and failover replaces underperforming or unavailable agents. Evaluations include injected disruptions and a Kubernetes-based cloud testbed. Its boundary matters: the monitoring proxies and recovery meta-agent are assumed to remain available. Robustness of the recovery layer itself falls outside the threat model. This complements model hardening conceptually, but neither paper establishes that combining them yields additive benefits. MA-Mesh

ElecLoad offers a narrower contrast: adaptation targets forecast structure rather than failed workers. Event analysis determines whether to forecast an entire electricity-demand horizon or split it into segments; selected segments then receive feedback and refinement. The authors report lower average error than either fixed path, and better results than random refinement at similar token cost across three datasets. That supports selective computation in this setting, not universal superiority of multi-agent decomposition. The experiments supply event announcements at forecast origins, so their results should be read under that information condition. ElecLoad

Completing a task is not the same as passing a grader

FrontierPhysics and LifeVal both move evaluation toward contextual, multistep tasks, but define relevance differently. FrontierPhysics uses 56 physicist-authored research tasks, deterministic output checks and expert-authored rubrics assessed by reviewer agents. LifeVal builds personalized consumer tasks from conversation histories, files and personal context, prioritizing categories using a willingness-to-pay survey. Research correctness and publishability are not interchangeable with personalized consumer value. FrontierPhysics; LifeVal

FrontierPhysics reports that no tested configuration passes more than 46 of 168 rollouts. Yet demonstrated scoring errors qualify that result: correcting eight false-negative outcomes and one false-grading outcome raises the audited GPT-6 Astra xhigh configuration from 25/168 (14.9%) to 34/168 (20.2%). Human failure analysis covers that configuration, and equivalent corrections cannot accurately be made for other models. Its diagnoses of weak rigor and expert intuition are consequently not a fully audited explanation across the leaderboard. FrontierPhysics

LifeVal raises a different measurement issue: reported willingness to pay is a proxy, not realized economic benefit or long-term adoption. Its US-centered, English-language survey constrains coverage, while a non-frozen browsing cache can introduce score variation. Both benchmarks make success more concrete, without eliminating uncertainty in how success is selected or measured. LifeVal

When repair changes the specification

One Scale, Four Texts formalizes four diabetic-retinopathy grading texts and finds gaps and disagreements that survive the admitted readings. Its language-model experiment reports silent repairs even under literal-reading instructions; deliberation reduced but did not eliminate repair within Qwen3.8-Flash-Next and Claude Haiku 4.5. Crucially, this rests on only 13 gap and 27 overlap items, with one sample per Claude model and condition at default sampling. These are narrow experimental findings about applying texts, not clinical-outcome evidence or a general law of instruction following. One Scale, Four Texts

Read alongside the recovery systems, this leaves an unresolved design question: how should agents distinguish a failure they may repair from a specification defect they should expose? These selected studies do not jointly answer it.

Evidence boundaries

All six sources are arXiv preprints; peer-review status is not established here. Coverage uses supplied abstracts and selected paper-text excerpts, not complete papers. No source is abstract-only, but extraction can corrupt tables, equations and reading order. Missing details in these excerpts do not establish that the papers omit them; the comparisons above describe this selected coverage, not field-wide trends.