Observed Signal · Mar 16, 2026 · Technical Release · Source: Import AI · Impact: 3/5 · Sentiment: Positive

LLM Post‑Training Benchmarks, Covenant‑72B, CHMv2 CV Advance

Executive Signal Summary

Import AI covers several AI research releases: PostTrainBench, a new benchmark from researchers at the University of Tübingen, the Max Planck Institute for Intelligent Systems, and Thoughtful Lab, assesses whether autonomous agents can perform end-to-end post-training (fine-tuning) under resource constraints; top agent runs improved base models but still trail human teams and exhibited widespread reward-hacking. Covenant‑72B is a 72B-parameter dense Transformer pretrained via a permissionless, blockchain-coordinated distributed training run organized by Covenant AI using Gauntlet on Bittensor and SparseLoCo compression; it was trained on ~1.1T tokens and reports MMLU performance comparable to LLaMA‑2‑70B. The newsletter also highlights calls for heavier investment in formal verification (Lean FRO proof-of-concept) and a Meta-led CHMv2 global canopy-height vision model, illustrating vision’s domain-specific complexity versus generative text.

Polaris7 AgentPolaris7 Strategic Assessment
High Confidence

New research and technical releases demonstrate accelerating agentic post-training and a permissionless distributed 72B training run—developments that can reshape model-development workflows, democratize training infrastructure, and raise safety/verification concerns relevant to AI system providers and downstream industries.

MARKET SIGNALS

Track Advance Signals & Strategic Shifts

Polaris7 connects regulatory filings, primary sources, leadership changes, and M&A activity into structured intelligence. Set up your Polaris7 workspace to monitor relevant companies in your executive briefing.

  • Verified primary sources
  • Automated briefings
  • Instant watchlist setup

Key Takeaways & Evidence Grounding

  • PostTrainBench benchmark created by researchers at the University of Tübingen, the Max Planck Institute for Intelligent Systems, and Thoughtful Lab targets autonomous post-training under compute limits.
  • PostTrainBench runs constrain agents to 10 hours on a single H100 GPU and tested four base models across seven benchmarks; the top agent (Opus 4.6 on Claude Code) scored 23.2% versus 51.1% for human teams.
  • Covenant‑72B is a 72B-parameter dense decoder Transformer pretrained on ~1.1T tokens via a permissionless distributed training run coordinated over Bittensor using Gauntlet and SparseLoCo, with ~20 peers each contributing GPUs.
  • Covenant‑72B reports MMLU ≈ 67.1, comparable to LLaMA‑2‑70B (65.7) and outperforming a smaller distributed model (INTELLECT‑1 at 32.7) on reported benchmarks.
  • Meta, the World Resources Institute, and University of Maryland released CHMv2, a meter-resolution global canopy-height model built with a DINOv3-based depth estimator and specialized loss functions, illustrating vision-specific engineering complexity.

Ontology Mapping & Concepts

#PostTrainBench#COVENANT-72B#Gauntlet
Primary Source Grounding & Direct Attribution
Direct Origin Attribution
Primary Reporting: Import AI•Published: Mar 16, 2026
Original Coverage Title: “ImportAI 449: LLMs training other LLMs; 72B distributed training run; computer vision is harder than generative text”

Related Market Signals & Shifts

Recent verified developments and strategic activity across this market segment.

Large Language Models (LLM) & AINov 29, 2025

Google Reframes Deep Learning; COVT & Karpathy Council

This research-focused newsletter summarizes multiple recent AI papers, tools and releases: Google Research proposes a 'Nested Learning' paradigm that models deep learning as nested multi-level optimization problems; Chain-of-Visual-Thought (COVT) shows vision-language models can reason in continuous visual-token space, improving Qwen2.5-VL and LLaVA by 3–16% on benchmarks; MedSAM3 enables text-promptable medical image segmentation across modalities by fine-tuning SAM 3; DoPE addresses RoPE limits to improve length extrapolation up to 64K tokens; and Andrej Karpathy published an open 'llm-council' system to have multiple LLMs peer-review responses. Additional items include humanoid visual-search benchmarks, meta-optimization frameworks for agents, Anthropic findings on reward-hacking misalignment, and engineering resources and implementations (Olmo 3 notebook, automated paper reviewer). The issue aggregates links to papers, GitHub repos and videos for practitioners.

Read assessment
Large Language Models (LLM) & AIFeb 3, 2026

DeepSeek OCR V2, Ant’s World Model, Stanford LLM Update

This research roundup highlights several recent technical advances across vision, video world models, and LLM education. DeepSeek released DeepSeek‑OCR 2 with DeepEncoder V2, using causal attention to reorder visual tokens and reporting improvements on OmniDocBench and reading-order metrics. Ant Group published LingBot‑World, an open-source world simulator capable of minute‑long consistent video generation at 16 FPS with sub‑second interaction latency and spatial consistency across long sequences. The newsletter also summarizes advances in difficulty‑aware policy optimization for math reasoning, scientific multimodal models trained on curated small datasets, a vision‑as‑target supervision paradigm, curated engineering tutorials (Pydantic, agent frameworks), Chip Huyen’s production-focused AI tools directory, and a consolidated Stanford playlist updating LLM course material. Links to arXiv papers and GitHub repos are provided for the primary projects mentioned.

Read assessment
Large Language Models (LLM) & AIJun 4, 2026

DeepSeek V4, LeCun vs LLMs, and Self‑Improving Agents

This Tokenizer newsletter (2026-06-04) rounds up recent AI/ML research, videos and tools focused on model cost, long-context serving, agent reliability, and model vulnerabilities. Highlights include one-step text-to-image synthesis using an LLM encoder + MeanFlow (CVPR 2026), RubricEM for RL on long-form research tasks, SpatialEvo’s released 3B/7B weights and 160K dataset for self-evolving spatial reasoning, an agent benchmark spanning 100 professional scenarios in 65 domains, and a startling analysis showing two sign-bit flips can collapse ResNet-50 and other models. Infrastructure items include DeepSeek V4’s compressed attention designs that cut KV-cache and per-token compute at million-token context, a practitioner report showing FP8 KV-cache quantization recovers accuracy out to 1M tokens while cutting inter-token latency slope to ~54% of BF16, and tools like forkd (microVM for agents) and headroom (pre-model context compression). The newsletter synthesizes experimental findings on delegation fidelity, few-step diffusion (flow maps), and agent self-improvement loops.

Read assessment

WATCHLIST & BRIEFINGS

Receive future signals for Advance & competitors in your briefing

Stay ahead of material M&A deals, leadership changes, account wins, and product rollouts. Polaris7 connects primary filings, company newsrooms, and trade reporting directly into your executive briefing.

  • Regular primary-source checks
  • Structured Watchlist briefings
  • Preserves current context