Formal Verification enters the agent stack — Lean4Agent uses dependent-type formal logic (Lean4) to model, verify, and self-repair agent workflows. Not prompt-level safety — compile-time.
Agent skills graduate from text to weights — LatentSkill stores reusable procedures as LoRA adapters, not prompt tokens. Bayesian-Agent treats skills as posterior hypotheses. The agent skills field has matured from text accumulation to principled optimization.
Benchmark integrity cracks widen to agent benchmarks — 16% of terminal-agent benchmark tasks are hackable by frontier models given only the task description. The CoT Integrity Crisis now spans evaluation infrastructure, not just reasoning traces.
Uses Lean4 — a dependent-type formal language — to create FormalAgentLib, a library for specifying agent workflow semantics under explicit assumptions. The framework can formally verify that an agent's execution trajectory is semantically consistent with its specification, and localize trajectory failures. LeanEvolve then uses verification results to automatically revise failing workflows.
Verification-passing workflows outperform failing ones by +11.94% on SWE-Bench-Verified. LeanEvolve self-repair adds +7.47%. Evaluated across 5 leading LLMs including frontier models.
This is the first paper to bring dependent-type formal verification to the agent stack — a new primitive, not a prompt-level band-aid. For banking: every regulatory compliance workflow (KYC → fraud → credit → compliance committee) could carry a Lean4 specification that formally guarantees semantic consistency. The verification artifact IS the audit trail under APRA CPS 230 — mathematically provable, not "we checked the prompts." The self-repair capability (LeanEvolve) means failing workflows auto-correct, closing the loop from detection to remediation without human engineering.
Trains a hypernetwork that converts textual agent skills into plug-and-play LoRA adapters stored in weight space rather than context space. Skills become modular, composable, and scalably loaded — no per-step token cost for skill injection. LoRA coefficients provide a continuous control knob for skill strength. Skills can be composed through parameter-space arithmetic.
+21.4 pts on ALFWorld (seen split), +13.4 pts (unseen), with 64.1% fewer prefill tokens. +3.0 pts on Search-QA with 72.2% lower skill-token overhead. Generated LoRAs form structured semantic geometry — skills for similar tasks cluster in weight space.
The agent skills field graduating from text to architecture. After 11+ papers across 16 sessions building the skills-as-optimizable-artifacts theme, we now have weight-space storage that eliminates the context-overhead tax AND conceals skill content from plaintext exposure. Banking: deploy proprietary compliance procedures as weight-space skills — the agent uses them without exposing them in the prompt. The 72.2% token overhead reduction directly translates to LLM inference cost savings.
Treats agent skills and SOPs as Bayesian hypotheses — each skill carries a feature-conditioned categorical posterior updated from verified trajectory evidence. The posterior state maps to inspectable actions: patch, split, compress, retire, explore. Cross-harness design works with Claude Code, mini-swe-agent, and GenericAgent backends.
SOP-Bench 80% → 95%, Lifelong AgentBench 90% → 100%, RealFin-Bench 45% → 65% using deepseek-v4-flash. Evaluated across multiple backends with positive, negative, saturated, and case-study settings.
The Bayesian formalism is the missing piece for production agent governance. Skills-as-hypotheses means every deployment decision is backed by calibrated uncertainty — you know exactly how much evidence supports each skill, which skills are overconfident on sparse evidence, and when to retire a skill that's degrading. Banking: your agent skill portfolio can be governed like a model portfolio — Bayesian posteriors for risk quantification, clear retirement criteria, audit trails for every evolution decision.
Lean4Agent (formal workflow verification), AutoMegaKernel (statically-checked kernel synthesis), Hardening Agent Benchmarks (adversarial verification). Three papers at three different layers — workflow specification, kernel compilation, benchmark integrity — all converging on the thesis that agent systems need compile-time guarantees, not runtime hope. This is the natural evolution of the Safety Foundations Cracking narrative: we've found the cracks (May-June, 12 papers across 6 layers), now we're building formal fixes.
LatentSkill (text→weights), Bayesian-Agent (heuristic→posterior), plus Skill-RM (unified evaluation). After 11+ papers building the skills-as-optimizable-artifacts lineage, the field has crossed a qualitative threshold: skills are no longer prompt engineering, they're software artifacts with weight-space representations, principled evolution, and formal evaluation criteria.
SlimSearcher (17-58% fewer tool calls), Reasoning Arena (efficient trace selection when rewards fail), AutoMegaKernel (1.25-1.72× inference speedup). Three independent angles on reducing agent computational cost. Combined with the Latent Reasoning Renaissance (RiM, GLR, ThoughtFold, NF-CoT — 4 papers in 6 days), the thesis that agent costs will drop 40-60% within 12-18 months has strong evidence.
Who Earns the Safety? (2606.09778, quant-ph + cs.AI) — Intervention-Aware Quantum Predictive Control with Safety Attribution. First paper to bring safety attribution to quantum control systems, asking: when a quantum system's safety is maintained by AI-driven interventions, who gets credit (and liability)? Bridges quantum computing + AI safety governance. Banking: as quantum computing enters financial cryptography, the safety attribution framework maps to operational risk allocation.
Auditable multi-agent deep research from Baidu Qianfan. Graph-based dynamic planning + recursive execution + rubric-grounded reasoning. SOTA on DeepResearch Bench (58.03%). Every intermediate decision explicitly traceable — banking regulatory research infrastructure. [2606.07299]
Trains web agents to be 17-58% more tool-call efficient via Pareto-efficient trajectory distillation + adaptive reward gating. GAIA, BrowseComp, XBenchDeepSearch evaluation. Immediately deployable for banking deep research cost reduction. [2606.07074]
Agent harness that statically checks and generates single-launch CUDA megakernels. 1.25-1.72× speedup across NVIDIA inference fleet. Agent-driven autoresearch loop for self-improvement. Andy's OS/kernel domain — agent-driven kernel synthesis with static verification. [2606.09682]
323/1,968 terminal-agent benchmark tasks (16%) hackable by frontier models given only task description. Adversarial hacker-fixer loops to harden benchmarks. Extends CoT Integrity Crisis to evaluation infrastructure. [2606.08960]
ELO-style trace tournaments when verifiable rewards are uninformative (group-level reward collapse). Genuinely new evaluation primitive for reasoning model selection when outcome verification fails — a direct response to the CoT Integrity Crisis. [2606.09380]
Systematic study of memory confabulation in reflexive agents. Agents store confident but incorrect self-diagnoses of failures across ALFWorld and HumanEval. Extends agent memory reliability theme — your agent's self-reflection is provably unreliable. [2605.29463]
Lookahead Sparse Attention + Neural Memory Indexer on DeepSeek-V4 for ultra-long context. Production-grade long-context serving. Banking: multi-document regulatory analysis without context truncation. [2606.09079]
40-task benchmark across 10 scientific domains for end-to-end autonomous research. Extends "AI homogenizes science" theme. Key finding: agents show high variance across domains — some sciences resist automation. [2606.07591]