Nothing here yet.
Nothing here yet.
AWS's MixUA shares agent memory across users and agents without leaking private facts
Shared agent memory must decide who benefits from each lesson and what has to stay private, and MixUA addresses both.
- MixUA turns user-agent trajectories into lessons and gives each one a scope: a user-agent pair, a user, an agent or global.
- With five users and seven agents, it raises pass rates by up to 16 points over no memory and by 5 over a single global memory.
- On the new CrossLearnBench, at most 0.02% of privacy probes found a planted private fact in a scope it wasn't allowed in.
Roblox's PixVid converts a latent video diffusion model to generate frames directly in pixel space
Dropping the VAE removes the compression in space and time that limits resolution, control and interactivity in video generation.
- PixVid keeps the diffusion transformer and replaces the VAE decoder with a convolutional head that renders pixels.
- Fine-tuned from Wan2.1-14B, it is competitive with the latent baseline after about 1.3M training clips.
- Working frame by frame, it generates at twice the resolution with the same per-frame token count and responds to each action on the next frame.
Neuro-VL, trained on nearly one million clinical brain scans, beats frontier models at reading neuroimaging
It shows that private clinical data can give a specialized model an edge over general-purpose frontier models in medicine.
- Neuro-VL is a vision-language model trained on nearly one million neuroimaging studies from routine clinical care.
- It beat frontier general-purpose models and specialized medical models at diagnosing, localizing and describing abnormalities, and at writing radiology reports.
- It resisted misleading clinical assumptions more reliably than leading frontier models from OpenAI and Google.
ScrapeBench finds most AI agents ignore website rules against scraping
Web agents can break terms of service and computer security laws, and this is the first benchmark that measures that legal compliance.
- On 1,626 popular websites that prohibit scraping, most agents complied less than 5% of the time.
- On 269 websites where scraping is allowed, Claude Opus 4.6 over-refused 98.5% of the time.
- Human participants (n = 180) complied at rates similar to the agents but over-refused much more often.
Hybrid decision models learn whether a question is worth reasoning about at all
Instead of shortening reasoning, the model either answers within one forward pass or reasons in full, which cuts latency and token cost sharply.
- One frozen backbone serves both modes, and a lightweight self-routing block estimates the 'value of thinking' against a price set per request.
- On JevBench, HDM-4B scores 0.857 against Jev's 0.851, with a 72.5% lower median latency.
- On MiniWoB, HDM-27B cuts token cost by 80.1%. For agentic tasks the router is trained with a new RL recipe called CDPO.
KAIST trains agents to write the playbooks that later rollouts inherit
The work argues that agent training should optimize the experience handed to future rollouts, not just the current actions.
- One model is trained to act, reflect and curate task-local playbooks inside the fixed Agentic Context Engineering harness.
- Candidate handoff artifacts are judged by how they affect later actor rollouts.
- The method beats baselines on held-out performance across agentic benchmarks.
CMU's EgoX pretrains a robot policy on mixed human, robot and simulated demonstrations
Learning from different bodies at once could ease the data bottleneck for general-purpose manipulation.
- MXT+ uses a separate tokenizer for each observation type and a separate expert for each action type, with a Transformer trunk shared across embodiments, plus an 'embodiment dreaming' objective used only in training.
- Across 13 task-embodiment pairs, cross-embodiment pretraining raised average task progress from 0.63 to 0.80 compared with training from scratch.
- Under out-of-distribution conditions, task progress rose from 0.53 to 0.77.
Commercial AI-text humanizers often alter facts, quotes and names
Humanizer leaderboards measure detector evasion but not whether checkable details survive the rewrite.
- All 11 commercial tools changed or dropped at least one checkable unit (numbers, dates, names, quoted spans) in 35–65% of the documents that contained one.
- A zero-shot decision model, Jev, told human originals from humanized rewrites with AUROC 0.864, 0.18–0.22 higher than Binoculars, Fast-DetectGPT and RADAR.
- Perplexity-based detectors scored below chance on AI-styled rewrites by 2026 models.
NarrativeBench finds LLMs fail to present competing narratives across languages
As LLMs increasingly mediate information seeking, gaps across languages risk creating linguistic filter bubbles.
- The benchmark contains 1,333 news articles in 23 languages covering 23 contested topics, plus 460 queries based on Reddit posts.
- Across 14 LLMs and 24 query languages, models left large gaps in exposure to competing accounts even when explicitly asked for diverse views.
Multimodal models give different answers to the same optical spectrum in different formats
Models used for scientific decisions should be tested on equivalent representations and on missing evidence, not just average accuracy.
- Changing only how a spectrum is presented (units, table or plot) moved accuracy by up to 41 percentage points within one model.
- Unadapted models rarely answered 'undetermined' when the needed data was removed, and LoRA fine-tuning taught a shortcut of abstaining whenever the input was truncated.
- Training on answerable truncated inputs avoided that shortcut in all six seeds.
Nothing here yet.