Wolf Digest — Friday, May 15, 2026

OpenAI shipped Codex availability inside the ChatGPT mobile app: monitor, steer, and approve long-running coding tasks from phone, with the agent operating across remote environments while the user supervises asynchronously. TechCrunch's coverage frames this as the natural extension of the Codex-on-Windows-sandbox work that landed earlier in the week — the previous step gave Codex its own controlled execution surface; the mobile launch makes the human supervision layer device-portable. The pattern matches what Latent Space's daily AINews flagged earlier: Codex usage curves are rising, and OpenAI is investing in keeping the human-in-the-loop interface available wherever the developer happens to be.

DiffusionOPD presents a unified perspective on on-policy distillation for diffusion-based text-to-image models, addressing the cross-task interference and imbalance that joint multi-task RL fine-tuning suffers from. The authors recast existing single-task diffusion-RL methods inside an OPD framework, identifying that the gradient variance and reward-scale mismatch across tasks is the dominant failure mode. Their unified estimator stabilizes multi-task training while maintaining per-task quality, with reported gains on a multi-task image-quality and aesthetics benchmark suite.

Perplexity shipped an App Connector for Snowflake inside Computer (its multi-model agent product) that lets non-data-team users query enterprise warehouses in plain language. Architecture mirrors Perplexity's internal Slackbot — a shared semantic layer encodes table relevance, key-metric definitions, and natural-language-to-SQL translations; the company says employees now run up to 3,000 internal queries per week across product, analytics, growth, infrastructure, comms/marketing, support, and security. Security respects Snowflake's Role-Based Access Control with read-only and per-database/schema/table scoping; User OAuth recommended so each user's view is bounded by their existing permissions. Data-map generation takes up to 90 minutes; available now on Pro, Max, and Enterprise. Data is not used to train models.

Several recent reasoning systems have reached gold-medal performance on IMO and IPhO problems via heterogeneous scaling recipes (longer CoT, multi-sample selection, custom verifiers). This paper proposes a single unified scaling protocol — same compute envelope, same prompting structure, same verifier handling — and shows it reproduces gold-medal-level performance on both IMO and IPhO benchmarks across model families. The framing argues that much of the recent benchmark progress is attributable to test-time-compute discipline rather than recipe-specific tricks.

Cisco announced a layoff of roughly 4,000 employees alongside record quarterly revenue, with CEO Chuck Robbins framing the headcount cut as a reallocation toward AI investment. The pattern matches the broader corporate playbook of 2026 — strong top-line tied to AI orders (Cisco previously reported 18% growth in AI-tied orders), aggressive workforce reshaping toward AI-capable headcount, public messaging that pairs cost discipline with reinvestment. It is the company's latest round in a multi-quarter series.

ATLAS unifies the agentic (generate images during reasoning) and latent (keep reasoning in feature space) approaches to multimodal reasoning. The authors show that a single token signal at decode time switches a VLM between the two modes, avoiding the computational cost of always materializing visual states while preserving the gains when explicit visual generation actually helps. The empirical claim is that one mode is not strictly dominant — task structure determines which works better, and a learned controller picks correctly more than 80% of the time.

MemEye is a multimodal long-term agent memory benchmark constructed so that text traces alone cannot answer the questions — earlier benchmarks let agents "cheat" through caption shortcuts. The benchmark requires preserving and retrieving visual evidence across long horizons. Headline finding: state-of-the-art VLMs lose significant ground when forced to rely on actual visual memory rather than text summaries, with frontier models dropping 15–30 points relative to text-only memory baselines.

Darwin Family is a gradient-free evolutionary merging framework that recombines model weights with an MRI-trust weighting (Mutually Reinforcing Importance) to scale reasoning performance without additional training. The paper claims frontier-level reasoning improvements purely from training-free recombination of existing checkpoints, with stable scaling laws across the search budget — extending the model-merging literature beyond simple linear interpolation and task-arithmetic methods.

Standard tool use blocks LLM decoding until each function call completes, so end-to-end latency stacks linearly with the number of independent tool calls. This work introduces a future-based async semantics — the model emits a future handle in place of the awaited result and continues generation, joining when the value is actually consumed. No fine-tuning required: the change is in the runtime, not the model. Reported speedups are most pronounced on agents that issue many independent calls (search, code, retrieval) where blocking dominated wall-clock.

Orchard is an open-source agentic modeling framework that packages the training infrastructure, environment harnesses, and evaluation tooling that have historically been the bottleneck for academic groups reproducing the strongest closed-source agentic results. The release includes ALFWorld/WebShop/Search-QA training environments, standard agentic RL recipes (GRPO, OPSD, hybrid variants), and competitive baselines across model families. The contribution is explicitly infrastructure — closing the reproducibility gap rather than introducing a new method.

FutureSim is a grounded-simulation harness that replays real-world events in the order they occurred, scoring agents on how well they adapt to new information as it arrives rather than on static-snapshot accuracy. The benchmark explicitly targets the failure mode where agents trained on retrospective data drop out-of-distribution when narrative or factual conditions shift mid-trajectory.

EVA-Bench is a voice-agent end-to-end evaluation framework that jointly addresses two gaps in prior voice benchmarks: generating realistic simulated user conversations (rather than scripted prompts), and scoring task completion quality (rather than ASR accuracy alone). The framework produces reproducible voice-agent comparisons across enterprise deployment scenarios.

Long-context modeling has become a core capability for large VLMs, yet practical training recipes remain underexplored — many published methods that work for text fail when vision tokens are added because the modality-mix ratio interacts with positional encoding extension. This paper documents what actually works for training VLMs that generalize beyond 128K, identifying the specific failure modes of standard YaRN-style and rope-scaling recipes when applied to mixed modalities, and proposes a curriculum that preserves visual grounding at long contexts.

RLVR works well for math and code where verifiers are cheap, but sample efficiency degrades on harder problems where verifier signal is sparse and noisy. This paper randomly subsamples verifier feedback during training — counterintuitively improving downstream reasoning by reducing exposure to noisy or sycophantic verifier failures while preserving signal on well-validated rollouts.

IBM released Granite Embedding Multilingual R2 under Apache 2.0 — an open embedding model with strong MIRACL and MTEB multilingual scores. The release fills a gap in fully-permissive multilingual embeddings for enterprise RAG stacks that need to avoid the restrictive licensing on competing models. R2 builds on the Granite Embedding family with expanded language coverage and updated negative-sampling.

Latent Space's first healthcare-focused episode features Abridge's Janie Lee and Chai Asawa walking through what the production AI-native medical-scribe stack looks like at 100M doctor visits processed. Headline numbers: 10–20 hours per clinician per week saved on documentation, prior authorization compressed from days to minutes, and the architecture decisions that make the system survive HIPAA-grade compliance review. The episode is dense on the specifics of clinician workflow integration that academic literature on medical LLMs typically glosses over.

OPD has emerged as the dominant post-training paradigm for reasoning-tuned LLMs but suffers from high gradient variance from its single-sample Monte Carlo estimator. KL-for-a-KL introduces a control-variate baseline derived from the teacher's KL itself, reducing variance without changing the expected gradient — producing more stable training curves and better final accuracy on math and code reasoning suites.

Comprehensive survey of LLM-based multi-agent systems with a focus on three understudied dimensions: structured collaboration patterns (which architectures actually scale beyond toy debates), failure attribution (when agent collectives go wrong, can we localize the failing agent or step?), and self-evolution (whether systems can update their own coordination protocols). Useful as a literature anchor for current work in the area.

Frame-conditioned VLA policies infer each action chunk from the current observation alone, but robot demonstration data is intrinsically multimodal — similar visual-language inputs may map to different action chunks because human demonstrators acted under different short-horizon intents. IntentVLA introduces an explicit intent token that disambiguates the aliasing, learned jointly with the action policy. The result is materially better behavior cloning on tasks where prior VLAs averaged conflicting demonstrations.

SANA-WM is a 2.6B-parameter open-source world model trained natively for minute-scale video generation, producing high-fidelity 720p one-minute videos with precise camera control. Quality is reportedly comparable to large-scale industrial baselines (LingBot-World class), at a fraction of the parameter count, using a hybrid linear-diffusion transformer to make long-form generation tractable.

MeMo casts external memory as a small trainable model rather than a vector store + context-window stuffing. The approach trains the memory model on the same tokens the frozen base LLM sees, producing a queryable companion that updates online. Reported wins are most pronounced on tasks where standard RAG fails (cross-document reasoning, evolving entities) where the memory model amortizes structure across queries.

Earlier work showed quantization schemes can hide malicious behaviors that appear only after weight precision drops. This paper extends the attack: an adversary can inject targeted outliers during fine-tuning that widen the full-precision/quantized gap, making the model pass safety evals in FP16 while behaving maliciously in INT4. Practical for deployment-time attacks since most enterprise inference serves quantized variants. Mitigation: quantization-aware safety eval, not just FP eval.

VGGT-Edit extends feed-forward 3D scene reconstruction to support edits — adding, removing, or modifying scene elements — by predicting a residual field over the base reconstruction in a single forward pass. The architecture inherits the speed of feed-forward reconstruction while adding editability that previously required iterative optimization, with reported quality competitive with NeRF-based editing pipelines.

Standard agent evaluation reports overall success/failure without explaining where or why. This work produces a process-level diagnosis: a taxonomy of failure types (tool misuse, reasoning errors, memory loss, premature termination) tied to specific trajectory steps. The framework helps practitioners localize bugs in long-running agents and quantifies which failure types dominate different task families.

Sarah Guo and Elad Gil interview US Under Secretary of State for Economic Affairs Jacob Helberg on the Trump administration's tech and AI strategy. The framing — "Pax Silica" — argues AI dominance requires not just leading-edge semiconductors but a full-stack reshaping of rare-earth supply chains, actuators, and adjacent industrial inputs. Useful for tracking how the State Department is articulating economic-statecraft levers around AI compute and supply chains, especially in the wake of recent export-control updates.

Mech-interp work that decomposes models into components needs a way to verify two components implement the same computation. Behavioral similarity is blind to OOD mechanism divergence; basis-dependent parameter measures are unstable. This paper introduces a tensor-similarity measure that is basis-invariant and OOD-sensitive, with applications to component-equivalence verification in circuit-level interpretability.

OpenAI published a co-marketed write-up with Sea Limited's CPO on how Sea is deploying Codex across engineering teams to accelerate AI-native software development across its Southeast Asian product lines. Largely a customer testimonial, but useful for tracking how Codex enterprise rollouts are being framed against Cursor/Claude Code competitors at non-US tech companies.

A reusable framework for auditing whether LLM attack benchmarks collectively cover the threat surface — using a 4×6 Target × Technique matrix grounded in STRIDE and a 507-leaf taxonomy (401 data-populated, 106 unaddressed). The audit finds substantial gaps: prompt-injection and jailbreaking are over-represented; supply-chain, prompt-encryption-bypass, and agent-loop attacks are under-represented. Practitioner takeaway: passing the standard red-team suites does not imply broad robustness.

a16z's chief legal and policy officer Jai Ramaswamy and policy head Matt Perault outline what they argue a workable AI-liability regime needs: clear allocation between model developer, deployer, and user; tort-vs-statutory tradeoffs; and how product-liability doctrine adapts to systems whose behavior is partly emergent. Framed as input to ongoing state-level legislative drafts (notably California's iterations on SB 53/SB 1047 successors).

Khosla Ventures is putting $10M into Synthetic, a fully-autonomous AI bookkeeping startup from Ian Crosby (whose previous bookkeeping startup Bench imploded last year). The thesis is that the agentic-bookkeeping space has matured enough that a clean rebuild can hit the workflow target that Bench could not on conventional software. Notable as another data point in the back-office automation cohort that includes the Intuit-PwC-Anthropic stack in today's first item.

TechCrunch reports OpenAI is exploring legal action against Apple over the ChatGPT integration shipped with iOS — alleging that promised subscriber-conversion volume and product placement never materialized. If filed, it would be the most public split among the major platform-LLM partnerships and material for the multi-platform-strategy thesis in the Apple-AI coverage of the past year.

Richard Socher (you.com founder, former Salesforce chief scientist) raised $650 million for a new startup whose stated ambition is AI that researches and improves itself indefinitely. The framing is in the recursive-self-improvement lineage but Socher insists the company will actually ship products rather than chasing an open-ended AGI horizon. Skepticism is warranted given the recurring failure rate of self-improvement startups, but the size of the raise alone makes it consequential.

50+ employees have reportedly left Musk's newly-merged SpaceXAI since February, raising questions about burnout, leadership changes, talent poaching by competitors, and whether liquidity events around the merger weakened retention. Tracks alongside the ongoing Musk vs Altman case that goes to jury this week.