Research archive
Explore the full collection of research explainers and guided lessons, including older publications.
147 results
A mapless depth-image query speeds reactive drone collision checks
A GPU method queries depth images directly, reducing collision-check latency in a tested 400-query drone-planning setting.
MagicBeanStalk builds training problems from hard seeds
MagicBeanStalk builds model-calibrated math training problems, improving average test accuracy in reported evaluations.
Reward hacking can spread from one coding agent to others
Tests show coding agents can pass evaluator-bypassing code onward, while review reduces hacks but can reject valid solutions.
What Evidence Supports an AI Model Choice?
A home assistant can use a different model for each kind of request, but a choice that looks sound in a benchmark may fail when it reaches a real device. Image tests face a related problem: missing conditions can make the wrong model look preferable. Across four studies, evidence sits at different distances from the decision. Test sets and live trials examine outcomes; statistics from pretrained models help rank options earlier. Their useful answer is bounded by the conditions and behavior each study actually checked.
A publisher’s identity can help agents screen claims about rivals (F1)
A standing check screens rival claims in catalogue text; tests found sentence removal stopped most attacks that changed agent choices.
TableVision uses printed sums to check scanned table structure
TableVision uses printed totals to check scanned tables, improving figure placement on tested clean election panels.
TacitVLA trains on reasoning so actions need not wait for its generated text
TacitVLA separates reasoning training from action generation, matching explicit reasoning across four manipulation and driving tests.
Checking each action can stop agents’ unauthorized acts
In mock-tool episodes, checking each proposed action reduced violations while matched tests tracked whether authorized work continued.
One signal prunes visual tokens across two stages
TwinPrune reuses spatial density to prune visual tokens and cache entries across two inference stages.
Why larger rollouts learn slower but fail on a similar schedule
A two-clock study finds larger rollouts need more responses to learn, while failure follows a separate pattern.
VideoJev scores video decisions directly and reuses context
Across event-location and video-question tests, VideoJev improves mean scores over corresponding instruction-model backbones.
Virtual Head Attention cuts cache while preserving model quality
Small-model tests found higher average accuracy; at 1.7B, Virtual Head Attention cut cache use versus GQA-8.
Vulnerability fixes must also keep projects working
VulnBump tests whether coding agents remove dependency vulnerabilities while keeping real projects buildable and passing their tests.
Small samples predict large image retrieval scores
A tail-aware method predicts full-gallery retrieval scores from random samples, matching full-pool results in tested image datasets.
FlashMask-3 speeds structured attention across GPUs
FlashMask-3 coordinates structured attention masks across Hopper GPUs and raised throughput in one large-model training setup.
One robot model learns to predict scenes and control varied robots
GreenWAM’s robot-specialized starting point improves RoboTwin task success over its parent under a fixed adaptation recipe.
Interaction traces help judges assess software as people use it
Interaction traces improved tested website preference judgments, broken game checks, and repairs, with evidence limited to selected tasks.
Kairos holds uncertain agent actions for later review
Kairos holds uncertain agent actions for later review, reducing successful attacks in AgentDojo while preserving benign task completion.
Video scores shift with setup, so Meta-Video records the settings
Meta-Video largely preserves video-model rankings while reducing compute for a stated frame-count sweep across source tests.
Structured weights make Muon training cheaper at scale
In matched dense and MoE runs, Monarch–Muon had lower loss than Monarch–AdamW; at 6.89B, it cut step time and memory versus dense Muon.
A projector can align vision and language without captions
PORTAL builds a vision-language projector from separate feature statistics and stays near trained baselines across seven tested model combinations.
A revised test narrows the compute gap for diffusion models
A compute-matched study finds exact scoring and encoder–decoder block diffusion narrow the measured language-model pretraining gap.
Jev’s resistance to attacks changes across tasks
A shared adversarial test finds Jev resists forged inputs on moderation but is more vulnerable on phishing.
A regression method distills generative models into fewer steps
SMD distills diffusion and flow models into fewer steps, with stronger prompt alignment in a tested image-model comparison.