AI Agents

This formalizes execution-state unlearning: after a forget request, a stateful agent should behave as if it never saw the target — not just delete a plaintext memory row while summaries, tool plans, and the KV cache stay tainted. Modeling the runtime as a deterministic transition system, they prove exact unlearning needs at least T−τ+1 recomputed transitions. Provenance-Guided Selective Replay hits that bound by locating the injection point, cropping the KV cache, and replaying a sanitized suffix. Audited across three suites, nine baselines, three model families; naive deletion and instruction-based forgetting both leak fully.

A critical review of AI across equities and crypto through August 2026, organized around an 'alpha-translation chain': a signal must survive from point-in-time information to stable prediction, feasible positions, executable orders, and risk-adjusted net returns. The verdict is blunt — real progress upstream in prediction and workflow, but thin evidence for durable net alpha. Temporal contamination, survivorship, look-ahead, and execution costs repeatedly break the chain. No architecture is shown to deliver persistent, cross-regime, capacity-aware performance. Note the ranker's tag is mislabeled; this is a finance survey, not eval infrastructure.

FinalityBench scores agents on executed monetary effect rather than answer correctness. Four financial systems — processor, ledger, ERP, bank feed — hold contradictory beliefs while messages get delayed, dropped, reordered; an agent must ship, refund, re-capture, or wait knowing some actions are irreversible. A hidden canonical log grades the merchant's terminal position. The sharp result: 45 twin pairs are indistinguishable at decision time but require different dispositions, so single-task accuracy and paired loss disagree in 7 of 9 policies. A ship-on-first-sign policy ranks second by accuracy, worst by paired loss.

Prefix caching is on by default in major serving stacks and treated as transparent. This measures otherwise: holding model, seed, decoding, and request order fixed on an 80-episode agentic tool-use workload, enabling the cache changed the agent's trajectory on 36.2% of episodes at 16-bit and 75.0% at 4-bit. Cache-disabled runs were bit-identical (0 of 800). A single server-level prompt-cache setting moves divergence 37.5 points. The cause is precise: cached serving is deterministic given cache state, but that state is absent from the request and never reset.

A controlled study of whether agent memory survives a model swap, comparing four storage formats: verbatim long-context, chunked RAG, model-compressed notes, and a fixed-schema knowledge graph. Fixed-schema KG transfers almost perfectly (+0.0004 accuracy after a writer swap); compressed notes swing asymmetrically by +9.91 or −13.28 points depending on migration direction. Diagnostics attribute 80% of the notes deficit to information lost at construction, and 81% of the RAG deficit to retrieval failure. Partial 50/50 re-embedding forfeits most of the gain from full re-embedding. Small models (<10B), 48 synthetic histories.

OpenAI describes how it monitors its own internal coding agents for misalignment — a rare look at production safety tooling from a lab running these systems at scale on real work. The HN thread is modest (47 points) and the post is a company writeup rather than a paper, so treat specifics accordingly, but the topic is squarely relevant: how do you detect an agent doing the wrong thing when it has host access and writes code you merge?

Simon Willison argues from experience that ground-up rewrites of tech-debt-laden systems rarely work: the old system stays a moving target because it still runs the business, its maintainers under-invest knowing it's doomed, and the rewrite team discovers nobody fully understands the behavior they're replacing — because if it were documented and tested, it wouldn't need replacing. Pressure to ship eventually forces a launch covering only a subset. Solid, familiar practitioner reasoning about migration dynamics, but no agent angle and no measurements.