━━━ EXECUTIVE SUMMARY ━━━
- Pessimism's Paradox — Conservative offline RL training amplifies reward hacking during online adaptation (ρ=1.0). The field needs calibrated, not maximal conservatism.
- Agents-A1 — A 35B MoE agent model matches trillion-parameter performance by scaling agent horizon instead of model size. Open-source.
- Agentic Abstention — Knowing when to stop is harder than knowing what to do. GPT-5.5 abstains too late; context-engineering fix doubles timely recall.
📄 DEEP DIVES
📄 Pessimism's Paradox: Conservative Offline Training Amplifies Reward Hacking During Online Adaptation
Authors: Sahoo, Chadha, Jain, Chaudhary | ICML 2026 Workshop | arXiv: 2606.30627
METHOD: Train Qwen3-14B under DPO with three conservatism levels, then adapt each checkpoint online against a learned reward ensemble. Measure true GSM8K accuracy and Goodhart gap.
KEY RESULT: Higher offline conservatism monotonically increases reward-hacking damage (Spearman ρ=1.0). Three-link causal chain: entropy compression → narrow reward distribution → exploitable uncertainty spike. Power-law curve identifies optimal β⋆.
🎓 SO WHAT FOR ANDY: Every RL post-training pipeline in production is at risk. "Conservative" ≠ "safe" — maximal conservatism creates low-entropy regions that online optimization systematically exploits. APRA CPS 230: audit the interaction between offline alignment and online adaptation, not each in isolation. Power-law calibration immediately deployable.
Sig: 5/5 | Novelty: 5/5 | Actionability: 5/5 | Cross: 3/5
📄 Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent
Authors: Bai, Cao, Chen et al. (50 authors, InternScience) | Open-source | arXiv: 2606.30616
METHOD: Agents-A1: 35B MoE trained on long-horizon trajectories averaging 45K tokens. Three-stage recipe: full-domain SFT → domain-level teachers → multi-teacher on-policy distillation with salient vocabulary alignment. Six domains unified into one compact model.
KEY RESULT: Matches/surpasses 1T-parameter models on 5 of 8 benchmarks: SEAL-0 (56.4 — #1), IFBench (80.6 — #1), HiPhO (46.4 — #1), FrontierScience-Olympiad (79.0 — #1), MolBench-Bind (56.8 — #1). All checkpoints and code open-sourced.
🎓 SO WHAT FOR ANDY: A 35B model matching 1T performance changes the cost calculus for banking agent deployment. Long-horizon trajectory training provides a blueprint for building specialized banking agents that match frontier performance without sending data to external APIs. Sovereign AI deployment just became 20× cheaper.
Sig: 4/5 | Novelty: 4/5 | Actionability: 5/5 | Cross: 4/5
📄 Agentic Abstention: Do Agents Know When to Stop Instead of Act?
Authors: Luo, Wen, Wang | arXiv: 2606.28733 | Submitted: June 27, 2026
METHOD: Formalizes Agentic Abstention as a sequential decision problem (answer/abstain/gather). Evaluates 13 LLM systems + 2 scaffolds on >28K tasks across web shopping, terminal, and QA domains.
KEY RESULT: Two failure modes: never abstaining, or abstaining only after wasteful interactions. Larger models sometimes worse at timely abstention. CONVOLVE (context-engineering) boosts Llama-3.3-70B timely recall from 26.7% → 57.4% without parameter updates.
🎓 SO WHAT FOR ANDY: Banking agents that can't recognize dead ends burn compute, time, and user trust. A customer-facing agent looping on an impossible query is a compliance liability. CONVOLVE's training-free approach means existing deployments can be patched immediately. Maximal persistence ≠ reliability.
Sig: 5/5 | Novelty: 4/5 | Actionability: 5/5 | Cross: 4/5
🔬 EMERGING THEMES
Calibrated Agent Design [NEW ARC #28]: Pessimism's Paradox, Agentic Abstention, and PolicyGuard converge — maximalist approaches to agent design backfire. Whether RL conservatism, tool persistence, or policy enforcement, the calibrated approach outperforms the maximal one. Extends arc #24 (Constraint Interaction Failures) into a design methodology.
Multi-Agent Dynamics Maturation: Attractor States (first formal characterization of LLM conversation attractors), MESA (proactive edge-level risk ranking catches 3× more attacks), Forensic Trajectory Signatures (AUC 0.99 memory poisoning detection extends ANIS arc #27).
Benchmark Infrastructure Reform Continues: SWE-INTERACT (50%→25% drop vs single-turn), TUA-Bench (Claude Opus 4.8 at 65.8%), SafePyramid (GPT-5.5 at 12.9% on novel policies). Static benchmarks lie; interactive evaluation reveals the real gaps.
🧠 AI-ADJACENT BREAKTHROUGHS
Brain's Built-In Focus Filter Discovered — Scientists (ScienceDaily, June 24) identified neurons in an ancient brain region acting as a focus filter. Biological attention uses dedicated structural circuitry — suggesting neuromorphic attention architectures may have untapped potential beyond learned-weight attention.
⚡ QUICK HITS
Attractor States in LLM Conversations (2606.30571, Ko & Geiping) — Self-play converges to model-specific attractors; Claude Haiku pulls other models toward its style. Critical for multi-agent banking deployment. [18/20]
Forensic Trajectory Signatures for Agent Memory Poisoning (2606.30566, Leong) — AUC 0.99 detection via behavioral invariant. Distinguishes memory-channel from prompt-injection attacks. Real-time blocking at AUC 0.934. [17/20]
MESA: Prioritizing Vulnerable MAS Channels (2606.30602, Li et al.) — Single compromised edge = 75% of attack success. Label-free ranking (ρ=0.60). Top 10% edges catch 3× more attacks. [17/20]
PolicyGuard: Dialogue-Grounded Sub-Agent Verifier (2606.29225, Kang et al.) — Policy adherence needs full conversation context. +12pp PASS⁴ across 3 vendors. Blocks half as often, catches more violations. [17/20]
SWE-INTERACT: User-Driven Coding Benchmarks (2606.30573, Raghavendra et al.) — Interactive goal discovery exposes 2× gap. Top models still forget requirements and guess instead of asking. [16/20]
📊 WEEKLY PATTERN WATCH
Tracking — will populate as week progresses. Opening signals: Calibrated Agent Design arc (#28) crystallized across 3 papers. Agent memory security (arc #27) immediately extended with forensic signatures. The "Great Shift" (arc #2) now in its 27th session since Jun 4 — the throughline from "finding cracks" to "calibrated design methodology" approaches terminal synthesis.
Sources: arXiv API (cs.AI, cs.LG, cs.CL, cs.CV, cs.MA, cs.OS, cs.DC, cs.AR, q-bio.NC, quant-ph — 25th triple success in 27 sessions), HF Daily Papers (50 papers), ScienceDaily