Use the right model for each task.
Bring OpenAI, Anthropic, Gemini, xAI, and Cursor into one research workspace.
Ascent: The platform for frontier artificial intelligence research.
world-models/latent_planning.ipynb⌘ K
main 91c3a0f
◫ latent_planning.ipynb ×
◫ ood_eval.ipynb ×
[12]
▶
import pandas as pd
import seaborn as sns
runs = pd.read_parquet("artifacts/ood_horizon_sweep.parquet")
summary = runs.groupby(["method", "horizon"])["success"] \
.agg(["mean", "sem"]).reset_index()
[13]
▶
sns.lineplot(data=summary, x="horizon", y="mean",
hue="method", marker="o")OOD planning successmean ± 95% CI · 8 seeds
latent rollouts recurrent baselineselected78.4%+12.6 pt
Comparison complete across 8 seeds
Mean +12.6 pp; 11 of 12 unseen environments improved.
Agents
Agents turn ideas into experiments.
Accelerate research by handing off work to Ascent, while you focus on making decisions. Compress weeks of research into hours.
Ascent
planning-circuits · main
Planning-circuit intervention study
Locate the smallest circuit required for multi-step planning. Use activation patching, random-head controls, and a locked behavioral suite.
Read notebooks/planning_probes.ipynb
Search src/model/transformer.py
Wrote plans/circuit-map.md
MDplans/circuit-map.md+84 lines
The plan fixes prompts, decoding, and token budget, then tests necessity and sufficiency independently.
Question
Which behavior should be the primary endpoint?
Skip
Add a follow-up...Plan⌄ Auto⌄ ↑
circuit-map.md ×interventions.py ×
Causal circuit discovery
Localize attention heads that causally affect multi-step planning.
Controls
Same prompts, decoding, token budget, and behavioral evaluation harness.
Search space
Layers × heads × intervention strength, replicated across three model sizes.
Success criteria
Ablation reduces accuracy >20 pp; rescue recovers >80% of the loss; FDR q<0.05.
3 tasks
○ Reproduce behavioral baseline
○ Run activation patching sweep
○ Replicate on held-out prompts
Autonomy
A hillclimbing machine.
Parallel agents explore competing directions, run fixed-budget experiments, and keep only measured improvements.
Ascent
world-models / overnight-07
Run 18 training
uv run train.py 04:11 / 05:00
best val_bpb0.9932↓ 0.0047
baseline0.9979experiment 01
kept6 / 1735% hit rate
best-so-farlower is better
Recent runscommit val_bpb verdict
17
add surprise-gated memory tokens7bd42f1
0.9932keep16
predict two future tokens jointly80c19aa
0.9941discard15
double the recurrent state2ea40d8
—crash
✓ gate passed→
run 5 min→
score holdout→keep if lower
Slack
Collaborate like it’s your colleague.
Launch experiments, ask follow-up questions, and generate results from your Slack workspace.
# ascent-labLaunch runs and review evidence
AscentApp
Started kras-g12d-cycle-03 with the requested holdout and off-target gates.
Scoring generated candidates68%
512 candidates24 model seeds8 GPU-h cap
Today
AscentApp
Computational screen complete; candidate SS-042 passed the locked gates. Not experimentally validated.
Candidate SS-04224 model seeds · completed in 2h 31m
Report ↗Predicted pDockQ
0.62→0.87+0.25
✓ 3↗ 23 replies
＋ Message #ascent-labAa @ ☺ ➤
Scientists
Meet your synthetic scientists.
Research agents work in parallel to test hypotheses, run investigations, and return the evidence behind every result.
Ascent
Scientists · 4 running
Ada-17Run wm-072 · awaiting review
completed in 01:42:18Hypothesis
Bidirectional latent rollouts let world models plan beyond their training horizon.
runs72
seeds complete8 / 8
OOD success78.4%
Best locked-eval scorehigher is better
Result ready for review
Mean +12.6 pp; 11 of 12 unseen environments improved.
Recent activitystatus
✓ seed-07 completed78.9%
✓ seed-06 completed77.8%
↗ reports/ood_summary.mdwritten
The new way to do frontier research.
Scale autonomous agents.
Launch fleets of agents that work in parallel on ambitious open-ended problems for hours or days.
Built around the way you work.
Connect your codebase, notebooks, data warehouses, and the enterprise tools your team already uses.
Contact us