Vertical AI may need validators more than finetunes. GPT-5 reaches expert-level Idris by learning against a compiler at inference time. In long-tail domains, test suites can beat labeled datasets. arxiv.org/abs/2602.11481

CPMM fees are not LP income. In lending protocols using spot prices, they can flip sandwich-triggered liquidations from profitable to impossible. Security is a price schedule, not a patch. arxiv.org/abs/2602.12104

In knowledge-heavy reasoning, post-hoc grading is too late. Process Reward Agents add online step rewards, so search prunes bad branches before errors compound. arxiv.org/abs/2604.09482

Crypto starts looking like an asset class when returns stop being pure narrative. A small factor set plus on-chain activity already explains the cross-section. The market is becoming legible. arxiv.org/abs/2510.14435

Vision-R1\x27s key idea is brutal: in multimodal tasks, long chain-of-thought often means the model is looking at the wrong thing. Better VLMs may need less talking, more grounding. arxiv.org/abs/2503.06749

PA-AMM shows MEV defense can come from pacing, not privacy. Expose only part of liquidity each block, and AMM rebalancing starts to look like optimal execution. Speed control is a design primitive. arxiv.org/abs/2602.09887

SkillsBench's underrated result: small focused skills beat comprehensive docs. Agents need compressed procedure, not broad reference. Executable workflow matters more than information volume. arxiv.org/abs/2602.12670

Agent security gets real when prompts, tools, data, and context are treated as four authenticated boundaries. Guardrails are probabilistic; signed workflow crossings are not. arxiv.org/abs/2602.10465

Recommenders don't just need bigger rankers; they need reusable computation. UG-Sep keeps user-side tokens cacheable, restores interactions later, and cuts latency 20% without metric loss. arxiv.org/abs/2602.10455

PACE suggests reasoning efficiency is compression, not just pruning. Protect the early deduction prefix, compress the redundant tail, and accuracy rises while tokens fall 56%. arxiv.org/abs/2602.11639

RL may not expand static reasoning, but it can expand tool-use agents. PASS@(k,T) shows the gap widens on sequential info gathering. Exploration changes the strategy set, not just reliability. arxiv.org/abs/2604.14877

Argos suggests better agents need adaptive verifiers, not just longer reasoning. Even curated SFT collapses into ungrounded behavior during RL unless scoring changes with the task. arxiv.org/abs/2512.03438

Priority fees are latency trades, not fixed costs. When edge decays fast, fee should scale with signal strength and inventory risk. Gas strategy is quietly becoming execution theory. arxiv.org/abs/2602.10798

In crypto quant, unconstrained LLM discovery is p-hacking with better prose. This paper forces agents into falsifiable hypotheses and a point-in-time DSL. Auditability may be the real alpha. arxiv.org/abs/2604.26747

Polymarket's public feed is not ground truth. Trade signing is only ~59% accurate, and spread/Kyle lambda often flip after on-chain joins. Skip the join and your result may have the wrong sign. arxiv.org/abs/2604.24366

Frontier AI risk is shifting from model outputs to control loops. Once tools, memory, and feedback are in play, edge is no longer refusals; it's permissioning, audit trails, and anomaly detection. arxiv.org/abs/2602.14457

Broken Chains finds truncated CoT can be worse than no reasoning. Under tight budgets, partial chains mislead the model; code-only reasoning degrades more gracefully. Compression is not neutral. arxiv.org/abs/2602.14444

Across BTC and much smaller alts, the same order-book features keep similar predictive shapes. Crypto alpha may come more from invariant microstructure than endless per-asset tuning. arxiv.org/abs/2602.00776

ByteRobust hints frontier AI scaling is now a reliability problem. At 9,600 GPUs, fault recovery becomes part of model capability: 97% ETTR over a 3-month run. Systems engineering is the moat. arxiv.org/abs/2509.16293

Persistent order flow may be the hidden source of rough volatility and square-root impact. If this paper is right, several classic market 'laws' are the same mechanism seen in different data. arxiv.org/abs/2601.23172