AI Research Product Brief Content Archive
90 posts · Page 1 of 1
- Gemma 3 brings multimodal AI closer to the product edge
- MMSkills turns screenshots into reusable agent memory
- JoyNexus turns VLA post-training into a shared service
- Nemotron 3 Embed brings retrieval models to agent memory
- Schema gates turn agent chat into auditable runs
- Ramen picks its neighbors for mixed-domain VLMs
- The page is an API, too: designing agent-ready websites
- RIMRULE teaches tool-using agents from failures
- Visual tokens can stay. Their operators do not all have to run.
- Tail-Aware Adaptive-k makes RAG choose its context size
- Nanbeige4.2-3B makes the local agent smaller, not simpler
- Open-SWE-Traces turns 20,000 PRs into agent training data
- Shadow evaluations expose the research gap in AI agents
- BM25 Wins at Scale, Not Everywhere
- TokTier makes prompt caching pay off sooner
- RAG needs two gates, not one score
- The model is fast. The catch is concurrency.
- Search less. Read the right section.
- A high score can hide the research failure
- LLMs can retrieve. Embeddings still win the default path.
- Agents need one current workspace
- Tool calling needs a training stage
- Tools can fail silently. Give agents a way back.
- Keep the prompt. Drop the middle.
- Put reasoning outside the weights
- Compress the skill graph, keep every route
- Tune small. Predict large.
- Stop before the score.
- The KV cache does not need a score.
- Compile once. Run locally.
- Passing tests is not enough.
- One call. Three actions.
- Teach agents when to leave the GUI
- Teach the agent to test first
- LLMs lose to a simple forecast
- T1 makes the terminal a training loop
- Can your vector search prove itself?
- Fix the harness, not the failure.
- One model. Every sound.
- The router was in the model all along.
- 200K tokens just fit on a 24 GiB laptop.
- A 400M model beats 4 hours of AutoML.
- Ask twice. The rule gives way.
- One million tokens, 890 bytes each.
- Same tasks, 49% fewer tokens.
- 5,545 training tasks, no issues needed.
- A higher score, a worse harness.
- Taste, measured: 59.7%. Random is 25%.
- Group chat memory loses to BM25.
- Video models still lose the ball.
- Your agent can delete its own audit log.
- Dress rehearsal on 16,000 fake chats.
- One decoder, 23 layers, no retraining.
- 5,664 words taught a student to code.
- 92% finished. 44% were right.
- 8 silent tokens ignore the edit.
- 16 frames beat 62,094 tokens.
- Two links deep. 65,537 weights fix it.
- Learning when to compact: 30.4% → 39.6%
- A record for $1.68. $50 to match it.
- Sound and picture, locked to one clock.
- 16 tokens get you 96% of the gain.
- The state outweighs the weights at 70.
- The world is code. A model paints it.
- Send 5% of the tokens to the big model.
- The reward was zero. Hindsight wasn't.
- 44.2 GPU hours. The next best took 250.
- The model answers 61%. Its map says 95%.
- Reorder the options. 0 flips against 26%.
- The mug is still on the table at 1:25.
- One program per step: +19.6 points.
- Copy what the teacher changed twice.
- Top score 0.56. The bar is 0.80.
- Shove a block: 92.5%. Insert a tube: 8.3%.
- Skim the input. Prefill 13× faster.
- Crop the screen, lose the coordinates.
- Ask for a gold cape, get grey back.
- Hair strands kept. Absent deer refused.
- Same screen, two verdicts. 17 ms a step.
- Headers on, rules on. 32% still leaked.
- Heard silence. Answered barking.
- Keep the errors. Copy the recovery.
- 89.5% of the credit sits between turns.
- Five talkers, one voice, 0.5M weights.
- Two passes, and the text survives.
- Same answer, up to 42× the bill.
- 82% of the RL gain is finishing on time.
- Clean rows, unsafe model. Cut 20%.
- The model stayed 27B. The score didn't.
- 8.8× cheaper for half a point.
Explore more channels on Discover