Wolf Digest — Wednesday, June 10, 2026

SWE-Explore argues that holistic resolved/unresolved benchmarks like SWE-Bench hide the sub-skills that make a coding agent work, and carves out repository exploration as a measurable capability on its own. Given a repo and an issue, an explorer must return a ranked list of relevant code regions under a fixed line budget; ground truth is derived from independent successful agent trajectories. It spans 848 issues across 10 languages and 203 open-source repositories, giving a fine-grained probe of context retrieval and code localization rather than end-to-end patch success.

The day after WWDC 2026, the analysis settled on what Apple actually conceded: a Siri rebuilt on Google's Gemini and Apple Foundation Models co-developed with Google, pitched with deliberately modest goals. Ben Thompson's “The iPhone's Last Stand” reads the keynote as Apple defending the device as the locus of personal computing even as the intelligence inside it is increasingly someone else's. A concrete artifact of that posture: Nvidia confirmed its Confidential Computing GPUs now run confidential inference inside Apple's Private Cloud Compute as PCC expands beyond Apple's own data centers onto Google Cloud.

Google released Gemini 3.5 Live Translate, a streaming speech-to-speech audio model that detects over seventy languages and produces near real-time translated speech preserving intonation, pacing and pitch. It is rolling out now in Google Translate on Android and iOS, with Google Meet getting a private preview for select Workspace customers and access via the Live API. The emphasis on prosody-preserving, low-latency speech-to-speech — rather than cascaded ASR then translation then TTS — is the technically interesting part, targeting natural-conversation latency.

WorldOlympiad diagnoses video-based world models along three axes that visual-quality benchmarks miss: physical faithfulness, geometric consistency, and interaction fidelity. The physical track uses object segmentation plus an MLLM-as-judge to check whether generated videos obey rules of mechanics, thermodynamics and material behavior; the geometry track reconstructs the generated video with Gaussian splatting to test 3D structural consistency across views; and an interaction track probes controllability over long horizons. It is a pointed counter to the tendency to score world models on how good the frames look rather than whether they model a coherent world.

Existing multi-state linear-attention methods improve long-context capacity but rely on fixed state-merging policies that cannot adapt to varying token importance, irreversibly obscuring critical tokens and accumulating error over long sequences. DLA proposes information-aware dynamic state merging, which sets state boundaries from token-level information variation rather than a fixed schedule, so important tokens are preserved instead of averaged away. It is a concrete step in the linear-attention line toward closing the long-context quality gap with full attention while keeping sub-quadratic cost.

Reinforcement learning with verifiable rewards is bottlenecked by weak reward contrast: trivially easy or hard prompts give low-variance feedback, and outcome-only rewards assign the same terminal credit to every decision in a multi-turn rollout. TRACE models each ReAct-style thought-action-observation turn as a distinct node and allocates rollout budget by prefix-level informativeness within a rollout, not just at the prompt level. It targets multi-turn agentic RL specifically, where the per-turn informativeness that prompt-level methods ignore is exactly where the learning signal lives.

ARM is a 7B discrete-representation autoregressive model that folds image understanding, generation and editing into a single next-token-prediction framework. The core is a discrete semantic visual tokenizer trained with joint objectives for semantic discriminability, language alignment and faithful reconstruction, mapping images into compact token sequences in a shared latent space. After large-scale text-and-image token pretraining, ARM applies RL to sharpen preference-aligned text-to-image generation and instruction-guided editing — an entry in the continuing push to make one autoregressive backbone do perception and generation together.

Role-Agent has a single LLM play both the agent and the environment to enable bootstrapped co-evolution without external feedback. In its World-In-Agent role the model predicts future states after each action, and the gap between predicted and actual state becomes a process reward that rewards environment-aware reasoning; in its Agent-In-World role it analyzes failed trajectories and retrieves harder tasks to train on. It is a self-play-flavored answer to the twin problems of sparse interaction feedback and static training environments that cap agent generalization.

Online RL methods for flow-matching image and video models (Flow-GRPO, CPS) cast denoising as an MDP and borrow PPO-style ratio clipping, but the authors argue clipping is structurally wrong for flow models: the single-sample probability ratio is a noisy estimate of true policy divergence, over-constraining some trajectory regions and under-constraining others. Flow-DPPO swaps ratio clipping for a divergence proximal constraint, exploiting the fact that each per-step flow policy is Gaussian to compute the divergence exactly. It is a cleaner trust region tailored to the geometry of flow-matching generators.

Keeping the full KV cache resident during decoding is the memory bottleneck for ultra-long-context serving. Lookahead Sparse Attention, built on the DeepSeek-V4 architecture, uses a neural memory indexer to proactively predict future context demands and keep only the query-critical KV chunks in GPU memory. The indexer is trained backbone-free as a standard dual-encoder with retrieval frameworks, so it never loads the full model into GPU memory during training — a “less is more” recipe that the authors say substantially raises serving throughput for long contexts.

As LLM agents start performing in-silico biology that once required trained biologists, ABC-Bench tries to measure the dual-use edge of that capability directly. The Agentic Bio-Capabilities Benchmark evaluates agents on both benign and dual-use tasks: writing code to drive liquid-handling robots, designing DNA fragments for in-vitro assembly, and — most pointedly — evading DNA-synthesis screening. By scoring the specific agentic skills that matter for biosecurity rather than abstract knowledge questions, it gives safety and policy work a concrete instrument, and it lands the same day Anthropic cited bio-uplift as a reason for Fable 5's biology guardrails.

One-step generative models accelerate sampling by learning deterministic flow maps from ODE dynamics, leaving open how to distill stochastic dynamics exactly. The Itô map is an any-step stochastic flow map that takes an intermediate state and a Brownian path and predicts future states in a single pass, yielding cheap differentiable access to posterior samples for inference-time control. Empirically it produces diverse, conditionally valid endpoint samples from fixed intermediate states and steers well on synthetic and image-generation tasks, establishing any-step SDE integration as a useful primitive for posterior sampling.

US Central Command deployed an autonomous Corsair maritime drone built by Saronic to find and recover two soldiers stranded near the Strait of Hormuz after their Army AH-64 Apache crashed during a patrol, a CENTCOM spokesperson confirmed. Defense One calls it an apparent first — an uncrewed surface vessel executing a personnel rescue of a downed helicopter crew at sea. The mission is a concrete data point on autonomous maritime systems moving from surveillance into time-critical, human-in-the-loop operations, and it comes amid surging US-Iran tensions in the region.

EEVEE is presented as the first multi-dataset test-time prompt-learning framework for LLM agents, aimed at the realistic case where inputs arrive as mixed streams from many datasets, domains and task distributions. To avoid cross-dataset interference, a router partitions incoming inputs into task clusters and assigns each a suitable prompt configuration, optimized through a router-prompt co-evolution that interleaves router and prompt learning to handle their mutual dependency. It reports improved robustness on real-world-style task streams where single-dataset prompt methods break down.

KV-cache compression methods either degrade quality, cost too much to compress a single long prompt, or require the input to fit the target context window and break modern inference engines. This work revisits encoder-decoder compression — mapping a long token sequence to a shorter sequence of latent embeddings a decoder consumes — and, via an architecture search, makes it competitive with KV-cache compression on the accuracy-efficiency frontier for the first time, while staying compatible with production serving.

After two days of testing at Marine Corps Air Station Yuma, the Pentagon's counter-drone task force approved SkyValor — a “detect and defeat” counter-UAS system from CACI International — for use across the military. Officials said it is capable of long-range targeting and round-the-clock automated sensing against unmanned aerial threats, validated against targets at varying ranges, elevations and flight paths. It is a marker of how quickly autonomous sensing-and-engagement systems are being fielded for base and border air defense.

On-policy distillation is increasingly used to improve LLM reasoning, but its training dynamics are poorly understood. Using parameter-space diagnostics, the authors place OPD in a relaxed off-principal regime: relative to supervised fine-tuning it touches fewer weights and avoids principal directions more strongly, while relative to RL with verifiable rewards it stays less tightly constrained. OPD also shows “subspace locking,” rapidly entering a narrow low-dimensional update channel — constraining training to that early subspace preserves OPD performance but wrecks SFT, suggesting the methods occupy genuinely different optimization geometries.

Most GUI benchmarks test general-purpose software on short-horizon tasks, leaving open whether agents can drive domain-specific professional applications to complete economically valuable work end to end. Workflow-GYM targets exactly that gap, with long-horizon GUI tasks in specialized professional software environments, and reports that state-of-the-art agents still struggle with the precise, multi-application, long-horizon workflows that real knowledge work requires — echoing Agents' Last Exam's finding from the GUI-operation side.

Extending LLMs to speech usually means cascaded ASR-LLM pipelines, heavy end-to-end speech-language training, or bridge/distillation adapters — each paying in latency, training cost, or sequential coupling. AuRA feeds the same speech to an ASR encoder (teacher) and a LoRA-adapted LLM (student) through a lightweight audio embedding layer, and uses layer-wise distillation to align the student's hidden states with the teacher's. The result internalizes audio encoding into the LLM itself as a cheap LoRA, avoiding both transcript-interface latency and full multimodal retraining.

Express is a tool for turning a non-causal attention approximation into a causal one with matching approximation guarantees. Combined with the Thinformer approximation it improves the best known causal-attention bounds — log^{3/2}(n)/s error with O(s) memory — and ships an I/O-aware Triton kernel showing substantial speedups over FlashAttention 2. The authors use it to attack four pipeline bottlenecks at once: long-context prefill, KV-cache compression, and memory- and compute-constrained long-form decoding.

Instruction-tuned LLMs are routinely post-trained into reasoning models for accuracy, usually without explicitly preserving alignment behaviors like safe refusal, bias avoidance and privacy protection. Auditing reasoning models produced via SFT, RL-based post-training and distillation against matched instruction-tuned baselines across six trustworthiness dimensions, the authors find the conversion is not behavior-preserving by default — reasoning training can quietly degrade safety and other alignment properties even as it lifts task accuracy.

Video world models that keep 3D consistency usually maintain explicit point-cloud memory in RGB space, which is expensive (repeated rendering and VAE encoding) and lossy (the pixel round-trip discards latent features). Mirage instead keeps a persistent 3D cache directly in the diffusion latent space, lifting latent tokens into 3D via depth-guided back-projection and querying it by synthesizing novel views through direct latent-space warping. Avoiding the pixel round-trip makes long-horizon spatial consistency both cheaper and higher fidelity.

In the US Marine Corps Aerial Logistics Connector program, Shield AI, Airbus US Space & Defense, L3Harris and Parry Labs completed a fourth autonomous flight test period on an H145 helicopter, the first with all four companies' systems fully integrated. During testing, Hivemind mission autonomy detected landing-zone obstacles in real time and maneuvered the helicopter to an alternate landing zone without crew commands — a step toward AI-piloted rotorcraft for contested logistics.

Multimodal LLMs increasingly act as the “brain” of embodied systems, but their grasp of physical tool use is largely untested. PhysTool-Bench is presented as the first physical-tool-use benchmark: 2,510 queries over 2,678 real-world tools spanning manufacturing, electrical work, agriculture and healthcare, evaluating whether a model can comprehend a scene, identify the right physical tool, and plan its use. It is a concrete probe of the perception-to-action gap that separates digital API tool use from real embodied competence.

To test document-automation capability, the authors adapt China's National Computer Rank Examination into a benchmark of 200 practical Word, Excel and PowerPoint tasks, each scored against a 100-point rubric with 7,118 machine-gradable criteria. Benchmarking seven frontier LLMs reveals stark gaps on the long-horizon planning, precise parameter configuration and multi-application integration that real office automation requires — a reminder that productivity-software fluency lags headline reasoning scores.

Evaluating six coding agents on four esoteric languages — with file editing, local execution and hidden-test grading — exposes capability differences that mainstream benchmarks like SWE-Bench Verified and Terminal-Bench 2.0 compress into a narrow band. The strongest agents (Claude Opus 4.6 and GPT-5.4 xhigh) often avoid writing the target language directly: on Brainfuck and Befunge-98 they write Python that generates and debugs the target-language code, a metaprogramming workaround that says as much about agent strategy as about raw language competence.

Customs and Border Protection signed a $71 million task order with GDIT for AI-powered autonomous surveillance towers to be deployed across the southern border, the latest call under an IDIQ contract worth up to $1.8 billion to modernize CBP's surveillance-tower network. The award extends the pattern of autonomous, AI-driven persistent surveillance becoming standard border infrastructure, with the attendant civil-liberties debate it carries.

Longevity scientist David Sinclair plans to launch human tests of an oral cellular-“reprogramming” drug as part of a $101 million XPrize Healthspan competition that rewards teams able to restore measurable immune, cognitive and muscle function to an earlier apparent age. The effort sits at the AI-for-biology frontier where reprogramming-factor discovery and aging-clock readouts increasingly lean on machine-learning models — and where the leap from mouse data to human “rejuvenation” claims remains scientifically fraught.

Google sharply cut the price of its budget AI subscription tier, a move TechCrunch reads as a warning shot in an emerging AI-subscription price war. The companion piece asks whether enterprises can “learn to love cheaper models”: if comparable workloads can be served by cheaper models without quality loss, the economics of AI shift materially — pressure that dovetails with the cost backlash Palantir is now exploiting against the frontier labs.

Meta signed its first AI data-center deal in India, partnering with Reliance on a 168-megawatt facility that will support Meta's global AI compute needs and can expand over time. The deal extends the geographic spread of frontier-AI infrastructure into India's fast-growing market and adds another node to the global compute-buildout map alongside this week's US-centric megaprojects.

The AI app-builder Lovable says it has surpassed $500 million in annualized run-rate revenue, with users creating about a million new projects a week and, the company claims, building real businesses and replacing internal software. The figure is another marker of how fast “vibe-coding” / natural-language app platforms are monetizing — a category Fable 5's one-shot app and game generation is poised to intensify.

A Lawfare conversation with Christoph Winter and Charlie Bullock works through how to govern transformative AI when capability is driven by scaling laws and the policy space is defined by what they call “radical optionality” — keeping many governance paths open under deep uncertainty about timelines and takeoff. It is a useful policy-side companion to a week whose technical news (Fable 5, the China blacklist) keeps outrunning the institutions meant to govern it.