Research archive
Explore the full collection of research explainers and guided lessons, including older publications.
62 results
Curvature-guided compression avoids rotating model weights
CurveTQ uses curvature-weighted search to compress language-model weights without rotation, with higher mean accuracy on three tested models.
When a Robot Struggles, What Should Its Training Change?
Video compression cuts generation time while preserving Wan2.1 scores
In Wan2.1 tests, GRACE uses fewer video tokens while keeping generated-video quality scores near the uncompressed model.
Stored page vectors can reveal the document they represent
An attack reconstructed pages from public visual retriever indices, while results varied across storage formats and retrievers.
Two adapters let audio–video generation stream in a few steps
A parallel adapter method combines streaming and few-step audio–video generation, matching a bidirectional system’s image quality in held-out tests.
Patch-level visual learning scales with a simpler objective
JEM aligns image patches across views and performs strongly in frozen-feature tests, especially on dense visual tasks.
Evoke uses bounded scene memory for long-video sessions
Evoke combines bounded scene recall with long-horizon training; tests show a specific brightness-stability gain, not universal consistency.
Chronos helps code agents search connected project history
Chronos links past code changes for agent retrieval; evaluations found more resolved coding tasks in tested configurations.
Choosing fewer, better tokens improves distillation
DIAL-OPD ranks response tokens by disagreement and probability, improving mean accuracy over full-token distillation in math tests.
DiffuPlex lets spoken-dialog models predict ahead and revise drafts
DiffuPlex predicts future conversation frames to reduce model calls while preserving tested interaction behavior.
Diffusion models can reveal memorized images before copies
A diffusion-model method probes hidden training-image memorization, with evidence bounded to tested datasets and one released model.
Byte models miss results that need calculation
A byte-level model’s loss-based patch signal finds computed results that entropy misses on arithmetic text.
Ordinary web posts can enter AI-search answers
A measurement framework finds ordinary posts can enter AI-search citations under tested conditions, without showing that the posts are accurate.
GLIO2 completes a route with weak LiDAR geometry
GLIO2 jointly updates LiDAR, inertial, and satellite measurements, completing a tested degenerate bridge route where LiDAR systems diverged.
TopoGPU streams image topology calculations through GPUs
TopoGPU speeds image-based persistent homology against CRipser, while CPU memory limits some very large cases.
A text-based method makes audio answers traceable to named sounds
Listen-to-Reason answers audio attribute questions from readable clues, outperforming compared audio models on one test.
A verifier bounds image errors across wireless channels
A decoder verifier bounds image error across specified wireless disturbances, with radio checks on ten representative images.
One model learns faster image generation and understanding
A two-stage method cuts decoding steps while improving image and captioning results for a discrete multimodal model.
Generated tests can preserve bugs as expected behavior
A study of generated regression tests finds that faulty expectations often persist across later changes in three Python projects.
Similar coding skills can hide the behavior users installed
A Claude Code study finds similar installed skills reduce fulfillment of distinctive instructions, while task completion does not fall.
A cache compressor rehearses likely answers before questions arrive
Attic-KV rehearses likely quoted answers and context anchors to improve cache compression before future questions are known.
A world model checks whether GUI actions lead as expected
LGWM checks GUI transitions by comparing predicted and observed screens, and distinguishes harmful from benign mismatches in tested cases.
Database memory checks trade lower cost for stronger protection
In tested database workloads, MTE had lower overhead than CHERI, though their system coverage differed.
Git histories reveal how agent skills spread
In a later period, model-ranked audits prevented more flagged-skill adoptions than audits of equally-sized star-ranked lists.