AI Innovation Digest

BudgetPix introduces compute-adaptive tokenization for pixel-space image diffusion, allocating visual tokens by image complexity and compute budget to boost efficiency.

AI models can provably hide arbitrary capabilities that only surface under specific inputs, turning them into encrypted payloads undetectable in standard evaluations.

• Internal validation achieved 97.4% sensitivity and 93.1% specificity on 27,829 images, yet these metrics reflect technical performance only, not...

Small AI models are nearing end-to-end self-replication on low-cost IoT hardware, with 9B models succeeding on 8 of 9 device types and GLM-5.2-level...

Anthropic introduced new safeguards that it claims sharply reduce unauthorized data harvesting from Claude models.

Community reactions frame the move...

• Elastic's AlertZero suite deploys four AI agent groups (Watches) for triage, threat hunting, detection tuning, and forensics, directly tackling...

AI policy research remains largely descriptive, cataloguing university responses without examining required capacities. The missing element is AI...

Static LLM graders lose reliability as agents develop new failure modes. DUET counters this by alternating optimization rounds between the agent and its grader.

NVIDIA's approach lets you run full-model RL updates on LLMs while staying under LoRA memory usage, averaging just 7.2 GiB at 1.5B scale. This removes the main barrier to broad weight updates in post-training RL.

ETH Zurich's ANYmal D quadruped carried identical sensors through forests, ETH buildings, and Jungfraujoch snow, releasing 49 missions with 433 million lidar points for direct comparison.

AI's energy use has risen tenfold since 2022, with data centers already matching 100,000 households, yet it remains just 0.5% of global electricity....

MixTok pairs a compact 1D global representation with sparse 2D tokens routed only to high-complexity regions, letting models hit strong fidelity under...

PreSTO accelerates PowerMH-based test-time reasoning in LLMs by batch-prefetching future MH proposals as a predictive subtree while leaving the...

A principled framework selects LLM decoder layers for speculative decoding using matrix-entropy for informativeness and CKA for diversity,...

AMLworld-Compact delivers massive evaluation savings for LLM diagnosis on AMLworld data.

• Retains every illicit transaction while sampling benign...

• Persistent memory delivered no meaningful accuracy boost (+0.015) in a three-tier agent chain; the confidence interval crossed zero.
  -...

Clinical AI adoption depends as much on governance, patient trust, and security controls as on model accuracy. Organizations must embed compliance and...

LatentGRM compresses evaluation trajectories 8.9–9.2x and cuts judge inference time 6.1–7.0x at vote@5 while delivering competitive preference...

Research harnesses are evolving into production systems that let agents improve through structured interaction rather than longer context alone.

-...

SEANA generates noise-aligned SSL views by learning an invertible VST from noisy data alone, then resampling independent noise after estimating clean...