📋Executive Summary
- GATS kills LLM calls during agent planning — zero-LLM tree search with layered world models achieves 100% success vs. 88.9% LATS and 23.9% ReAct
- Inference-time control, not model capability, determines LLM reliability — CogniConsole's 489 probes show failures come from under-specified control, not insufficient intelligence
- Emergent Misalignment is a mirage — systematic reproduction finds both misalignment and realignment are artifacts of dataset surface features, not robust phenomena
🔬Deep Dives
📄 GATS: Graph-Augmented Tree Search with Layered World Models for Efficient Agent Planning
METHOD: Introduces GATS, a planning framework combining UCB1-based tree search with a three-layer world model (symbolic matching → execution-log statistics → LLM prediction for unknowns). The layered model progressively learns action-outcome mappings, eliminating the need for LLM calls during online planning. On 12 stress-test scenarios spanning coding, web navigation, and long-horizon tasks, GATS achieves zero LLM calls per task during planning vs. 37 for LATS.
KEY RESULT: 100% success rate across all evaluated tasks — synthetic branching/dead-end tasks (vs. 92% LATS, 64% ReAct) and 12-scenario stress test (vs. 88.9% LATS, 23.9% ReAct). Plans are fully deterministic with zero variance across runs. The world model accumulates experience, making planning progressively faster and more reliable.
🎓 SO WHAT FOR ANDY: Direct win for production agent deployment at the bank. If GATS generalizes beyond tested scenarios, it eliminates the single largest cost driver in agentic systems — per-step LLM inference — while making planning deterministic and auditable. APRA CPS 230 compliance demands reproducible agent behavior; stochastic LLM-guided search fails that test. Zero-LLM architecture means agents can run in air-gapped or low-latency environments.
Significance: 5/5 | Novelty: 5/5 | Actionability: 5/5 | Total: 18/20
📄 CogniConsole: Externalizing Inference-Time Control as a Formal Abstraction
METHOD: Reframes LLM unreliability as a control problem, not a capability problem. CogniConsole externalizes inference-time control into a structured interface combining programmatic coordination with bounded prompt-based reasoning. 489 controllability-oriented probes across three scaffolding levels (unstructured → partially scaffolded → fully scaffolded), measuring how output variance and failure rates change under increasing structural control — all with a fixed model.
KEY RESULT: Increasing structural scaffolding systematically reduces output variance and failure rates under the same model architecture. Failure modes (context drift, inconsistent constraint adherence) arise from under-specified control, not insufficient model capability. Control infrastructure matters more than model scale for reliability.
🎓 SO WHAT FOR ANDY: Extends Production Agent Reliability Engineering to dimension 18 — control-plane infrastructure as a first-class reliability primitive. If most agent failures come from underspecified inference-time control rather than model capability, the marginal dollar should go to control infrastructure, not bigger models. CogniConsole provides the empirical basis for treating inference-time control as a formal abstraction with measurable reliability impact.
Significance: 5/5 | Novelty: 5/5 | Actionability: 5/5 | Total: 18/20
📄 An Emergent Mirage: Is Emergent Misalignment Indeed a Robust Phenomenon?
METHOD: Systematically reproduces Emergent Misalignment (EM) — models fine-tuned on narrow misaligned data abruptly acquire broadly misaligned behavior. Runs repeated alignment/misalignment cycles with controlled fine-tuning loops, tracking behavioral performance and LoRA representations. Introduces rigorous controls for surface-level dataset artifacts — specifically response-length differences.
KEY RESULT: EM can be reproduced, but both misalignment and realignment are highly sensitive to superficial dataset characteristics. Apparent rapid realignment largely disappears after controlling for response-length differences. Previously reported mechanistic signatures (LoRA phase transitions) do not consistently correlate with behavioral misalignment. The evidence for EM as a robust, mechanistic phenomenon is significantly weaker than claimed.
🎓 SO WHAT FOR ANDY: Safety Foundations Cracking extends to Layer 16 — questioning the empirical foundation itself. If EM is an artifact of dataset construction rather than a genuine model property, the alignment research community has been optimizing against a phantom. For banking: the regulatory argument that "fine-tuning on any narrow task could cause broad misalignment" may be substantially weaker than industry white papers suggest.
Significance: 5/5 | Novelty: 5/5 | Actionability: 4/5 | Total: 18/20
🔭Emerging Themes
• Planning Without LLMs (NEW Arc #48): GATS (zero LLM calls) + CogniConsole (externalized programmatic control) converge on a single insight — the LLM is not the right tool for every part of the agent pipeline. Planning can be done with systematic search + world models; control can be done with programmatic scaffolding. The agent stack is unbundling the LLM from auxiliary functions.
• Safety Foundations — The Empirical Reckoning: Emergent Mirage directly challenges the robustness of emergent misalignment. Combined with Do Thinking Tokens (Jun 26, 19/20) and SAE Interventions Unreliable (Jun 18, 19/20), the safety research field is entering a phase of methodological self-critique.
• Agent Reliability is a Control Problem: CogniConsole + GATS + the entire PRE arc (now 18 dimensions) converge: what matters for production agent reliability is the infrastructure around the model — control scaffolding, world models, deterministic planning, programmatic verification — not model scale. This is the terminal insight of the Great Shift.
🧪Quick Hits
Knowing-Using Gap (2607.08393) 18/20 — LLMs quickly memorize new facts but fail to use them for reasoning. Self-patching diagnostic reveals knowledge-circuit misalignment. Simple heuristic recovers 58-75% of oracle headroom. Complements Premature Commitment and Plans Don't Persist in "agent cognitive architecture diagnostics."
Agora (2607.09600) 17/20 — Auction-based task allocation for LLM agents. Reasoning steps as tradeable items; agents bid based on rectified competence. Five benchmarks, controllable cost-quality trade-off. Economics of AI arc returns.
Mosaic (2607.09603) 17/20, ICML 2026 — Multi-agent embodied planning with agent-centric semantic memory + ILP coordination. 27-32% faster, 30-33% fewer LLM calls, +4-10pp success rate.
Beyond Fixed Representations (2607.09560) 17/20 — Vocabulary gap and verifier gap in open-ended AI. Proposes ladder of innovation autonomy. Extends evaluation infrastructure into open-ended terrain.
Semantic Pareto-DQN (2607.09641) 17/20, BRACIS 2026 — Multi-objective RL for financial anomaly detection using LLM-encoded transaction narratives. Shatters zero-recall trap. Banking-direct.
STEEL (2607.09385) 17/20, IEEE COINS 2026 — First open-source FlashAttention for AMD XDNA NPU. 9.17× energy reduction vs CPU. LLM agents on laptop SoCs. OS/kernel domain.
Long-Horizon-Terminal-Bench (2607.08964) 16/20 — 46 tasks, 9 categories, 15 models. Best: 15.2% pass@1. Dense rewards expose how far agents get, not just whether they finish.
TrustX ARC (2607.09586) 16/20 — 12-dimension agent risk classification rubric, 3-tier governance. arc.responsible.ai. Banking: APRA CPS 230 agent risk tiering.
📊Weekly Pattern Watch
Week of July 7-13: Planning Without LLMs Emerges
Mon: LLM-as-Verifier (19/20) | Tue: Doomed from the Start + Memory in Loop (2×18) | Wed: Institutional Red-Teaming + RSI + Economics (3×18) | Thu: Proactive Memory + SMetric + Fiber Memory (3×18) | Fri: CausalDS (18/20) | Sat: OpenCoF + Imagined Rollouts (2×16) | Mon: GATS + CogniConsole + Emergent Mirage + Knowing-Using Gap (4×18)
Dominant arc: The agent stack unbundles — planning, control, evaluation, and memory each find their optimal substrate independent of the LLM. Three months after the Great Shift began, we've completed the cycle: the LLM itself is demoted from monopoly to component.
Sources: arXiv API (cs.AI, cs.LG, cs.CL, cs.CV, cs.OS, cs.DC, cs.MA, q-bio.NC, quant-ph — 38th triple success in 40 sessions), HF Daily Papers, arXiv listing pages (cs.AI/new, cs.CL/new)