Research Log · Monday, October 12, 2026

The harness is becoming the learnable object

Daily working notes on autonomous-vehicle AI and frontier AI research.

Top developments

SimVTA breaks the logs-only supervision bottleneck for driving world models. The OpenDriveLab method post-trains driving world-action models with GRPO and flow-matching on simulation-rendered trajectory supervision in 3D Gaussian splatting scenes, lifting NAVSIM v1 PDMS from 86.2 to 91.3 (arXiv 2610.13781). Simulated trajectory supervision as a post-training recipe is the data-engine idea worth studying here.

Four independent groups converged on training the harness, not the weights. CMU's Harness Learning applies RL to harness code and reports a 4B proposer beating a 35B teacher (paper); NVIDIA's VERA co-evolves model and harness across 9,000+ verifiable environments (paper); HERMES uses resident LLM Dev-Primitives to lift migration success from 6.5% to 31.0% (paper); and Sakana AI's MASS runs multi-agent self-supervision with no external verifier (paper). When four labs land on the same idea independently, the leverage has moved: the harness, not the model weights, is where the gains are right now.

Meta's Sharpening Tax warns that RL post-training shrinks solution coverage. Base models solve at large K what post-trained versions never solve again (paper), a direct caution for anyone leaning on best-of-N or verification strategies; PTGS is the proposed mitigation. Read this before your next post-training run.

Frontier labs

GPT-6 Sol and Luna shipped with generated UI; the safety regressions matter more. Responses now generate UI components on the fly (announcement), but the more important signal is in the system card: OpenAI admitted statistically significant regressions versus GPT-5.6 on self-harm evals, plus gore and sexual-content evals for Luna. Capability up, safety down, stated plainly.

The US restored UK AISI pre-release model access after a 17-day pause. During the gap, Anthropic shipped Claude Mythos 5.1 US-only (report). Voluntary cross-border eval access looks fragile: labs will re-geofence releases the moment politics intrudes.

OpenAI's Decisions API entered public beta, confirming the decision-model category. Structured verdicts instead of prose at $0.10 per 1M tokens (announcement), the second launch in a week after Cloudflare's Clef-omni. This category is now a trend, not an experiment.

DeepMind's EmbeddingGemma 2 is an open embedding model worth noting. A 740M-parameter open-weight multimodal embedder under Apache 2.0, 768 dimensions with Matryoshka truncation (announcement), belatedly captured from its October 6 launch.

AV / embodied

SplitJEPA takes the theory-first route to world models. A latent world model with invariant and variant subspaces and no reconstruction decoder, with formal identifiability under Gaussian predictive dynamics (arXiv 2610.12349). The theory-first world model papers keep getting stronger.

Splat2Act is the consumer phone-to-policy pipeline. A phone scan becomes a 3D Gaussian splatting twin, which yields 68k sim-oracle demos, which train a personalized pi0.5 policy, going from 0% zero-shot to 62.5% pickup (arXiv 2610.13945). The consumer version of the data-engine story.

Chronos Policy bakes physical time into robot policies. Explicit time-aware rotary position embeddings tolerate async sensors and variable latency (arXiv 2610.13929). Deployment-relevant: real robots have messy timing.

Noise Space Learning makes frozen VLAs cheap to adapt. Frozen flow-matching vision-language-action models are adapted by learning the noise distribution instead of the weights (arXiv 2610.13943). Cheap adaptation of big VLAs is a practical win.

Snow makes LiDAR detection easier, says a new IROS 2026 result. Snow accumulation boosts point counts, so weather-severity bins mislead without object-level accounting (arXiv 2610.13891). An evaluation-hygiene result that should change how people bin their weather evals.

Waymo went to the bond market and into Detroit. Its first debt financing, a $5 billion term loan led by PIMCO, Blackstone, and Sixth Street, coincides with the start of fully driverless operations in Detroit. NVIDIA's Ali Kani publicly gave Waymo the lead over Tesla, a notable shot from a key supplier. Bond-market money at this scale says the unit-economics story is credible.

TomTom and DeepRoute.ai signed an MoU for L2+ map priors. Report

At AutoSens, OEMs are rethinking monolithic end-to-end driving over chip and memory costs. Report The architecture pendulum keeps swinging.

Agents / tools / models

CorpusMap makes agentic retrieval cheaper and better, a rare combination. Entity-graph navigation for agentic retrieval improves quality by 6.4 to 11.7 points while cutting tokens 34 to 57%, with no LLM build cost (paper).

VeriHarness has the model's own generator verify its rollouts. With disagreement resolution and a consensus challenger, and about 26K rollouts released (paper). Self-verification harnesses are converging with the harness-learning trend above.

Agent Plasticity measured test-time learning in dollars. The gains-per-dollar metric for test-time learning puts paid GPT-6 Sol as the most efficient learner (paper), a useful efficiency lens on the just-run-it-longer instinct.

MIMESIS is a 9B user simulator that transfers. RL against it generalizes across unseen user simulators (paper). Transferable simulators are the bottleneck for agent training.

Sakana's Continuous Memory Machines put dual matrix-valued memories on continuous-time models. Beating LSTM, DNC, RMC, and CTM baselines (paper). Memory architectures are quietly having a moment.

NVIDIA's cross-model KV cache transfer cuts handoff cost 25x. Plus Cache-2-Cache (report). Concrete prefill savings for multi-model agent stacks, and the kind of infrastructure win that compounds.

Other notable

Anthropic's usage policy now forbids sustained and needless abusive or cruel behavior toward the model. TechCrunch reports the change alongside new election-interference language. Model-welfare norms are going mainstream in legal text: a cultural marker worth noting.

Claude Haiku 5.5 joined the GitHub Copilot model picker. Report A distribution win for Anthropic's small model.