- Program-as-Weights (19/20): A 4B compiler compiles fuzzy-function specs into 0.6B LoRA adapters that match 32B model quality — 1/50th the memory, 30 tokens/s on a MacBook. Foundation models become tool builders, not runtime problem-solvers.
- Mastermind (17/20): Strategy — not action trajectories — is the right learning unit for agentic software engineering. An independently trained planner boosts a frozen GPT-5.5 executor from 60% to 84.5% on vulnerability reproduction and transfers to other models without retraining.
- GRPO Identity (17/20): Three of the most popular RL-for-reasoning methods (GRPO, Dr.GRPO, DAPO) are proven to be three settings of one dial — the group standard deviation. A unanimous group teaches nothing; a split group teaches everything.
METHOD: Many everyday programming tasks — alerting on important log lines, repairing malformed JSON, ranking search results by intent — resist clean rule-based implementations and get outsourced to expensive LLM API calls. PAW introduces fuzzy-function programming: a 4B compiler model (trained on FuzzyBench, a 10M-example dataset) takes a natural-language spec and emits parameter-efficient adapters for a frozen 0.6B Qwen3 interpreter. The result is a compact, locally-executable neural artifact — invoked once per function definition, cheap for every subsequent call.
KEY RESULT: A 0.6B interpreter executing PAW programs matches direct prompting of Qwen3-32B while using ~1/50th the inference memory and running at 30 tokens/s on a MacBook M3. The paradigm reframes foundation models from per-input problem solvers into tool builders: the heavy model runs at compile time (definition); the lightweight artifact runs at application time, preserving locality, reproducibility, and cost efficiency.
🎓 SO WHAT FOR ANDY: This is a deployment paradigm shift for banking AI. Instead of routing every fuzzy classification/repair/extraction task through expensive API calls, a bank could compile domain-specific fuzzy functions once and deploy them as tiny local adapters — privacy-preserving, offline-capable, and auditable. The 30 tok/s on a MacBook means on-premise deployment is practical. For agentic systems, this enables compiling tool-calling heuristics, input validators, and policy checkers into lightweight artifacts that run inside the agent loop without API latency.
METHOD: LLM agents can execute repository-level vulnerability reproduction steps when the approach is correct, but they fail by choosing the wrong strategy. Mastermind is a dual-loop framework: a trainable planner learns reusable high-level vulnerability-reproduction strategies via SFT and milestone-based GRPO, while a frozen executor follows the strategy. Critically, the planner is trained independently of the executor — one planner can improve multiple different executor models without modifying their action-generation capabilities.
KEY RESULT: On CyberGym (260 train, 200 test), Mastermind with GPT-5.5 achieves 84.5% pass rate vs. 60% (open-book), 63% (Best-of-8), and 77% (iterative improvement). The same planner improves GPT-5.4 mini from 45%→60% and GLM 5.1 from 58.5%→71% — without any retraining. Strategy is compact enough to optimize, concrete enough to guide execution, and stable enough to reuse.
🎓 SO WHAT FOR ANDY: This directly applies to banking software security. A strategy planner trained on vulnerability reproduction patterns could be deployed across multiple agent executors — improving code review bots, security audit agents, and patch validation pipelines without per-model tuning. The cross-model transferability is the killer feature: train once on the best executor, deploy everywhere. The "strategy as learning unit" framing also applies to any agentic task where the approach matters more than the execution details — compliance checks, regulatory reviews, fraud investigation workflows.
METHOD: Three of the most popular methods for training LLMs to reason — GRPO, Dr.GRPO, and DAPO — look like three different tricks. This paper proves they are three operations on one number: the group standard deviation of correctness marks. For binary (right/wrong) rewards, the standard deviation equals the magnitude of the training update. GRPO divides by it, Dr.GRPO drops the division, and DAPO discards groups where it's zero — all three are operating on the same dial.
KEY RESULT: The "group-standard-deviation identity" reveals that a split group (half right, half wrong) provides maximum learning signal, while a unanimous group teaches nothing. This means problem selection and sampling budget can be optimized directly: only problems with real disagreement drive learning. Validated on Big-Math and controlled training runs. The identity also explains when and why these methods work or fail — it's the dial that decides where learning happens and how strongly.
🎓 SO WHAT FOR ANDY: This is the kind of theoretical clarity that saves months of trial-and-error in RL-based agent training. When fine-tuning agents for banking tasks, knowing that split-vote examples drive all the learning means you can: (1) curate training data for disagreement rather than difficulty, (2) allocate sampling budget to problems near the decision boundary, and (3) debug training failures by inspecting group variance, not just loss curves. The paper demystifies RLHF → a single number controls everything.
• ProtoPilot: Autonomous agent executes wet-lab protocols with 90.2% expert preference — AI crosses from in silico to in vitro. From the cs.AI listing page (Jul 3 submissions).
• CreativityNeuro: Weight steering boosts divergent thinking by 14 human percentile points — first demonstration that creative capacity can be dialed up without task-specific fine-tuning.
• Expander SAEs: Sparse autoencoders with 293× fewer decoder values than dense SAEs while retaining 84% of CE-loss recovery — mechanistic interpretability at practical scale.
• Conformal Policy Control: Finite-sample safety guarantees for RL policies via conformal prediction — production-ready statistical safety for agentic systems.
Week of July 6 (Monday start): First briefing of the week. Tracking these nascent threads:
• Compilation as Deployment: Program-as-Weights inaugurates a new capability frontier — large models as compilers, small artifacts as deployment units. Watch for follow-on work applying this to agents, vision, and multi-modal tasks.
• RL Theory Convergence: GRPO Identity + MIPU + DemoPSD all arrived in the same weekend. Is the RL-for-LLMs community collectively converging on the same theoretical foundation? Watch for synthesis papers.
• Agent Architecture Maturation: The decomposition trend (strategy/execution in Mastermind, state continuity in ElephantAgent) is accelerating. The agent stack is becoming a proper engineering discipline with separable concerns.
• Benchmark Saturation Crisis: Perf-Bench Audit adds to growing evidence that agent evaluation infrastructure is failing. 85%+ task saturation means leaderboards are measuring noise. The evaluation reform arc continues.