Wolf Digest — Wednesday, June 17, 2026

NVIDIA's Nemotron 3 Ultra technical report details a 550-billion-parameter (55B active) MoE hybrid Mamba-Attention model pretrained on 20 trillion tokens, context-extended to 1M, and post-trained with SFT, RLVR, and multi-teacher on-policy distillation. Key ingredients include LatentMoE, multi-token prediction, NVFP4 pretraining, and reasoning-budget control. Released open, it lands around 38 on Artificial Analysis's Intelligence Index. The combination of a hybrid linear-attention backbone with FP4 pretraining at this scale is the notable systems result, pushing open-weights efficiency on long context.

OpenAI burned through 3.7 billion dollars in the first quarter of 2026, more than half of its 5.7 billion in revenue, according to documents shared with shareholders. Both cash burn and revenue roughly tripled year-over-year, underscoring how monetization continues to lag demand. The company closed the quarter with more than 73 billion in cash and marketable securities, up from 40 billion at the end of December following its late-March funding round. The figures sharpen the question of how long even the best-capitalized labs can sustain frontier training and serving costs.

ChatGPT's share of AI-assistant usage fell below 50 percent for the first time, to 46.4 percent by the end of May, per Sensor Tower's State of AI 2026 report. It remains the leader with over 1.1 billion monthly users, ahead of Gemini (662 million, 27.7 percent) and Claude (245 million, 10.3 percent). First-half 2026 saw about 2.3 billion AI app downloads and over 4.2 billion dollars in spending, more than double the prior year; time spent roughly doubled to about 36 billion hours, with the top three apps commanding 89 percent of it. Claude leads on conversion at 13 percent paying.

Zone of Proximal Policy Optimization proposes steering a model's policy through teacher guidance injected in the prompt rather than through gradient updates. By placing a stronger teacher's hints in context and optimizing within the student's "zone of proximal development," the method aims to capture much of the benefit of on-policy RL without the cost and instability of weight updates. The paper reports gains on reasoning and benchmark suites, and was widely surfaced across arXiv categories and the daily-papers feeds, making it one of the day's most cross-referenced results.

Virginia-based Enabled Intelligence is adding more than 500,000 hours of curated drone footage from the war in Ukraine to its EView dataset library, the first Ukraine full-motion video in the collection. The footage is pre-labeled and validated for training across aerial object detection, vehicle classification, and ground-activity recognition, spanning electro-optical, SAR, infrared, and foreign-language audio. The company, founded in 2020, won a 2025 NGA data-labeling contract worth up to 708 million dollars over seven years that underpins the Maven program. CEO Peter Kant stressed the data is "real, not simulated," and available now to approved users in the US, Ukraine, and NATO nations.

Looped World Models apply weight-tied iterative refinement, looping a shared block multiple times, to world-model prediction, trading depth for repeated computation over a compact parameter set. The approach targets better long-horizon rollout consistency and sample efficiency without growing model size. Surfaced across cs.AI, cs.CL, cs.CV, and cs.LG plus the daily-papers feeds, it reflects a continuing thread of looped or recurrent-depth architectures applied to generative and predictive modeling.

Google DeepMind is partnering with the UK government on a Gemini-built prototype to halve decision times on householder planning applications, in support of a target of 1.5 million new homes by 2029. Householder applications make up nearly 70 percent of yearly planning cases. The tool consolidates data, surfaces local policies with citations, summarizes consultation feedback, and drafts assessments, with officers retaining final decisions and a full audit trail. It builds on the Gemini-based "Extract" tool, now offered to every council in England and estimated to save about 255 hours of manual work per council per year. National rollout is planned from 2027.

LoopCoder-v2 scales test-time computation for code generation by looping a single pass rather than sampling many independent completions, aiming for the accuracy of heavy test-time search at a fraction of the token budget. The method targets efficient inference-time scaling on coding benchmarks and was broadly cross-listed across agents, cs.AI, cs.LG, and evals feeds, part of the day's strong showing for efficient test-time-compute techniques.

The Ling-2.6 and Ring-2.6 technical report presents a model family for efficient agentic intelligence: Ling-2.6 optimized for instant, high-capability-per-token responses, and Ring-2.6 tuned for deeper reasoning and agentic workflows. Rather than training from scratch, the team upgrades the Ling-2.0 base through architectural-migration pretraining, reusing prior compute while changing the architecture. The report emphasizes low-latency serving alongside strong reasoning, a recurring theme in this generation of agent-oriented MoE releases.

This work proposes a unified autoregressive multimodal model with a shared context-visual tokenizer, letting a single token space serve both understanding and generation rather than maintaining separate visual encoders and decoders. The design aims to simplify multimodal architectures while preserving generation quality, and it was one of the most heavily cross-listed papers of the day, appearing across cs.CV, efficiency, generative-media, evals, and RL feeds alongside the daily-papers aggregators.

Ternary Mamba applies grouped quantization-aware training to reach W1.58A16 precision, ternary weights with 16-bit activations, on state-space models. By tailoring the QAT grouping to Mamba's recurrence, the method preserves accuracy while sharply cutting weight memory, targeting resource-constrained deployment of SSMs. It is a concrete step in extending extreme low-bit quantization, already explored for transformers, into the linear-recurrent regime.

"Learning from the Self-future" introduces on-policy self-distillation for diffusion language models, using a model's own later-step (future) predictions as targets to supervise earlier denoising steps. The approach improves diffusion-LM training without an external teacher, addressing the on-policy distribution mismatch that hampers dLLM optimization. It reflects growing momentum behind diffusion-style text generation and the post-training tricks needed to make it competitive.

OPD-Evolver builds a self-improving agent through on-policy distillation, having an agent generate and then distill from its own successful trajectories to evolve a more capable policy holistically rather than tuning isolated skills. The method targets compounding agentic improvement and was cross-listed across cs.CL, efficiency, and evals feeds, part of a cluster of on-policy-distillation work appearing this cycle.

The Qwen-RobotManip technical report argues that careful data-and-objective alignment is what unlocks scale for robotic manipulation policies, showing that aligned pretraining lets a manipulation model benefit from larger and more heterogeneous data. Part of a coordinated Qwen-RobotX release (alongside navigation and world-modeling reports), it reflects Alibaba's push to bring foundation-model recipes systematically into embodied control.

Variable-Width Transformers let the hidden width vary across layers or tokens rather than holding a fixed model dimension throughout, allocating capacity where it is most useful and trimming it elsewhere. The paper studies the accuracy-efficiency trade-offs of non-uniform width and reports favorable compute savings, contributing to the broader line of work on adaptive and heterogeneous transformer architectures.

GameCraft-Bench evaluates whether agents can build playable games end-to-end inside a real game engine, testing long-horizon planning, code generation, asset integration, and iterative debugging in a concrete, executable setting rather than on isolated snippets. The benchmark exposes large gaps between current agents and shippable game artifacts, adding to the growing set of execution-grounded agentic evaluations.

ACE-Ego-0 unifies egocentric human video and robot demonstration data for vision-language-action pretraining, aligning the two domains so a policy can learn manipulation priors from abundant human footage before transferring to robot embodiments. The work targets the data-scarcity bottleneck in VLA training and continues the human-to-robot transfer thread that several robotics-foundation efforts have pursued this year.

PseudoBench measures how agentic auto-research systems can manufacture and amplify pseudoscience, probing whether autonomous research agents, given a misleading premise, produce confident but spurious findings. By quantifying this failure mode, the benchmark surfaces a concrete misuse and reliability risk as auto-research agents proliferate, sitting at the intersection of evaluation and safety.

ERQA-Plus is a diagnostic benchmark for reasoning in embodied AI, decomposing embodied question-answering into targeted reasoning categories so failures can be attributed to specific capabilities rather than aggregate scores. It aims to give robotics and VLA researchers a sharper diagnostic signal than coarse success rates, and was cross-listed across robotics, evals, and robotic-autonomy feeds.

S4oP performs operator-level pruning of structured state-space models, removing whole SSM operators rather than individual weights to shrink models for resource-constrained deployment while keeping the recurrence structure intact. The operator-granularity approach is tailored to SSM internals and reports favorable accuracy retention at high sparsity, extending structured-pruning techniques into the state-space family.

This paper develops a function-space theory arguing that catastrophic forgetting in continual learning is predominantly low-rank: the functional changes that erase prior knowledge concentrate in a few directions. If the forgetting subspace is low-dimensional, it can be identified and protected cheaply, pointing toward lightweight continual-learning methods. The result offers a clean theoretical lens on a long-standing problem.

This position paper argues that prevailing coding benchmarks are misaligned with real agentic software engineering, rewarding isolated puzzle-solving while ignoring the multi-file, long-horizon, tool-using nature of actual development work. The authors call for benchmarks grounded in realistic repository-scale tasks and developer workflows, echoing a broader unease about whether current leaderboards measure what practitioners care about.

On Interconnects, Nathan Lambert hosts Finbarr Timbers for a detailed review of how frontier labs actually assemble post-training recipes, walking through the interplay of supervised fine-tuning, preference optimization, RL with verifiable rewards, and synthetic data. The conversation is a practitioner-level map of where the marginal gains in modern post-training come from and how the stages fit together, useful context for reading the day's many distillation and RL papers.

Snap unveiled a new pair of AR glasses priced at 2,195 dollars, pushing its long-running smart-glasses bet into a higher-end, AI-assistant-centric form factor. The launch lands amid a broader wave of AR-glasses and wearable-agent announcements, as platform companies race to put multimodal assistants into always-on eyewear. Price remains the obvious barrier to mainstream adoption.

Google's "Earth AI" work applies geospatial foundation models to nature restoration, moving from pixel-level satellite interpretation to actionable planning for where and how to restore degraded land. The research combines remote-sensing imagery with planning models to prioritize interventions, illustrating the continued push of large geospatial models into conservation and environmental decision support.

A FedScoop report argues that agentic AI is arriving across federal agencies faster than the governance guardrails meant to constrain it, as departments deploy autonomous, tool-using systems ahead of settled policy on oversight, auditing, and accountability. The piece captures a recurring tension as agentic deployments outpace the controls, complementing the day's Pentagon and benchmark-reliability stories.

NVIDIA detailed its XR AI stack for bringing multimodal agents to AR glasses, offloading perception and assistant workloads to enable hands-free, context-aware help through wearable displays. The announcement fits the week's cluster of AR-and-agent hardware news, framing eyewear as a near-term surface for always-available multimodal assistants.

DARPA is exploring technologies for tactically responsive space operations, seeking the ability to launch, reposition, and task spacecraft on operationally relevant timelines rather than over months of planning. The effort points to growing demand for autonomy and rapid decision-support in space mission management, part of the broader defense push toward faster, AI-assisted operational tempo.

A Two Minute Papers video walks a general audience through recent Anthropic interpretability findings on the internal mechanisms of Claude, popularizing results on features, circuits, and the sometimes counterintuitive ways the model represents concepts internally. While an explainer rather than primary research, it reflects continued public attention to mechanistic interpretability and the effort to make its findings legible beyond the field.