• LLMs won't scale their way to human-like social simulation — 85-model scaling study reveals compute improves opinion modeling but fails catastrophically on cognitive biases like risk aversion (Ziems, Hashimoto et al., Stanford)
• Multimodal AI overestimates shared understanding — VLMs conflate what both parties can see with what has been communicatively established, threatening collaborative agent reliability (Li et al., SIGDIAL 2026)
• Cross-domain transferability is the missing curriculum signal — gradient-geometry alignment between domains beats learnability-only curricula by 10% relative in multi-domain RLVR training (Yang, Schölkopf et al.)
METHOD 85 Qwen3-architecture models trained under fixed compute budgets (10¹⁸–10²⁰ FLOPs) plus 35 larger open-weight models up to 70B. Scaling laws applied to three social simulation domains: opinion modeling, behavioral simulation, and longitudinal forecasting.
KEY RESULT Opinion and behavioral simulation improve rapidly with scale — but human cognitive biases are entirely immune. Models from 0.5B to 8B parameters show zero improvement on risk-aversion calibration and correlated-reward learning, even with fine-tuning. Underrepresented populations scale more slowly; longitudinal forecasting weakly correlates with MMLU.
🎓 SO WHAT If you're using LLM agents to simulate customer behavior, stress-test regulatory scenarios, or model counterparty decision-making under APRA CPS 230 — scaling the model buys coverage breadth, not depth. The model's risk preferences won't converge to human norms just because it's bigger.
METHOD 13,077 annotated reference expressions from HCRC MapTask dialogues. VLMs evaluated on interpretation-matching under systematically controlled manipulations of dialogue context and map-information access — visual vs. textual vs. non-informative.
KEY RESULT Authentic map images improve overall performance but cause models to over-predict alignment — conflating what both participants could see with what has been communicatively established. Qwen3-VL-8B exhibits the strongest bias. The effect is driven by task-relevant map content, not the visual channel itself. Models rely on static referential cues rather than tracking how grounding unfolds through dialogue.
🎓 SO WHAT Multi-agent banking systems sharing visual workspaces (loan documents, dashboards, compliance screens) will silently assume alignment when none exists. Extends last session's "Asking Frontier" from when to ask to what has been understood.
METHOD TAC — a bandit-style online curriculum that selects training domains based on cross-domain transferability, not just local learnability. Repurposes signals already in the GRPO step: per-domain advantages for learnability, projected gradients for transferability via gradient-geometry alignment. <1% wall-clock overhead.
KEY RESULT +2.8 points (10% relative) over learnability-only bandit on 6-domain reasoning suite (Qwen3-1.7B, Llama3.2-3B). Removing the transferability term causes sharp degradation. On imbalanced training mixtures, learnability-only curricula over-commit to dominant domains; TAC remains robust.
🎓 SO WHAT Any multi-skill agent training pipeline at a bank should be measuring cross-skill transfer, not just per-skill improvement. The richest training signal is at the capability boundary, not the peak — and it's already in the gradient geometry you're computing.
• The Collaboration Frontier (NEW arc #33): Last session established that knowing when to ask beats raw capability. Today extends this to knowing what has been understood (Seeing Is Not Sharing) and knowing which domain's learning transfers (TAC). Three papers, three layers of collaborative intelligence: ask, ground, transfer. The progression mirrors human team dynamics — from tool use to shared mental models.
• The Compilation Paradigm (extends arc #7): PAW reframes LLMs from per-input problem-solvers to per-definition tool-builders — a 0.6B interpreter executing compiled neural adapters matches Qwen3-32B at 1/50th the memory. Extends: HyperTool → ToolPro → PAW. Each generation moves agent logic from runtime inference toward compile-time artifacts.
• Simulation Scaling ≠ Simulation Fidelity: The social simulation scaling study exposes a pattern we've now seen across safety, governance, and collaboration: scaling improves the measurable, misses the essential.
Week of Jun 30 – Jul 5:
Mon 30: Calibrated Agent Design — pessimism must be calibrated, not maximal
Tue 1: Safety as Geometric Structure — benign fine-tuning geometrically reverts safety (19/20)
Wed 2: Memory becomes the Agent Reliability Frontier — AutoMem 19/20, MemSyco-Bench 18/20
Thu 3: Iterative Attack Surface opens — 5× 18/20 Friday-level batch
Fri 4: The Asking Frontier — interaction strategy overtakes raw capability
Sat 5 (today): The Collaboration Frontier — from asking to grounding to transfer
The arc: Calibrated design → Safety geometry → Memory → Attack surface → Asking → Collaboration. A progression from internal model properties through external interaction.