Ascent: The platform for frontier artificial intelligence research.

world-models~/research⌘ K

Experiment Log Compute

Sessions

OOD horizon sweep41m Running · evals.ood_planning · seeds 0–7 Ready for review · +12.6 pp at h = 32

Activation patching2h Waiting · one question for you

Baseline replication4h Done · 8 of 8 seeds · 65.8%

New session

Files

configs

data

experiments

latent_rollouts.pyM

recurrent_baseline.py

notebooks

latent_planning.ipynb

ood_horizon.ipynb

results

MASTER_PLAN.md

README.md

latent_planning.ipynb latent_rollouts.py

[13]:

# rollouts vs. shortest path
latent = policies.load("runs/latent_h16/seed3")
viz.rollouts(latent, env="maze-l",
             horizons=[8, 16, 32], seed=24)
viz.divergence(latent, steps=32, samples=20);

Last executed at 2026-09-09 11:02:19 in 2.31s

[ ]:

recurrent = policies.load("runs/recurrent/seed3")
viz.rollouts(recurrent, env="maze-l",
             horizons=[8, 16, 32], seed=24)

Mode: Command Ln 3, Col 41 Python 3 (ipykernel) | Idle Trusted

OOD horizon sweep

latent_planning.ipynb

Run the locked OOD suite at horizons 4–32 over eight seeds. Compare latent rollouts with the recurrent baseline.

Running the sweep Thought for 6s

Editedexperiments/latent_rollouts.py+28−6

Ranuv run -m evals.ood_planning --seeds 0-738m41m 12s

Added2 cells to latent_planning.ipynb

All eight seeds finished. At horizon 32, latent rollouts reach 78.4% against 65.8% for the recurrent baseline (+12.6 pp, 95% CI +8.1 to +16.4). The gain grows with horizon; maze-xl is flat.

Horizon	Latent	Recurrent	Δ (pp)
8	71.2	70.4	+0.8
16	74.9	68.1	+6.8
32	78.4	65.8	+12.6

1 file changed+28−6UndoKeep all

Plan, search, run anything

Agent GPT-6 Astra

The Ascent workbench — live experiments, evidence, and the next hypothesis.

Agents

Agents turn ideas into experiments.

Give Ascent a research goal. It surveys the literature, develops an experimental plan, runs the work in parallel across nodes on models of any size, monitors every run, and returns results for you to inspect and build on.

Learn more

planning-circuits~/research

Experiment Log Compute

circuit-map.md interventions.py

planscircuit-map.mdDraft · agent, 2m ago

Causal circuit discovery

Smallest head circuit behind multi-step planning.

HypothesisLayers 7–9 carry the planning signal; ablating those heads should break multi-step success alone.

Candidate heads outlined · cluster at layers 7–9, heads 2–5

Success criteria

Ablation cuts planning accuracy by more than 20 pp
Rescue recovers more than 80% of the loss
FDR q < 0.05 across the candidate set

Method

Patching sweep, all heads, random-head controls
Necessity: ablate candidates, measure the drop
Sufficiency: rescue with candidate activations

Tasks

Reproduce the behavioural baseline

Run the activation-patching sweep

Replicate on held-out prompts

Planning circuit study

Find the smallest circuit behind multi-step planning. Activation patching with random-head controls.

Thought for 12s and ran 2 tools·48s

Readnotebooks/planning_probes.ipynb

Wroteplans/circuit-map.md+84

The plan is drafted. One decision before I spend compute:

Question Which behaviour should be the primary endpoint?

Maze success rate

Execution exact match

Both, Holm-correctedrecommended

SkipContinue

Plan, search, run anything

Plan Claude Fable 5.1

The agent reads the codebase, writes a reviewable plan, and asks before spending compute.

Autonomy

A hillclimbing machine.

Ascent orchestrates autonomous agent swarms that tirelessly explore competing directions, test hypotheses, and share what they learn. Give them a strong eval and a compute budget, and they return production-ready breakthroughs.

Meet Synthetic Scientists

lm-hillclimb~/research

Experiment Log Compute

overnight-07.md train.py eval.py

resultsovernight-07.mdAgent · just now

overnight-07

Synthetic Scientist on train.py · 5-minute budget per run · eval.py locked at 91c3a0f · baseline 0.9979

Run	Change	val_bpb	Verdict
18	Two-layer surprise gate	`0.9929`	keep
17	Surprise-gated memory tokens	`0.9932`	keep
16	Predict two future tokens jointly	`0.9941`	discard
15	Double the recurrent state	`—`	crash

Best so far

Best val_bpb after each run · lower is better · 18 runs, 8 kept

Synthetic Scientist · overnight-07

Improve val_bpb on the fixed five-minute budget. Edit only train.py; eval.py is locked. Keep a change only if the holdout score is lower, and log every run to results/overnight-07.md.

Run 16 · predict two future tokens jointly·discard 0.9941

Run 17 · surprise-gated memory tokens·keep 0.9932

Run 18 · two-layer surprise gate ·04:11keep 0.9929

Editedtrain.py+9−2

Ranuv run train.py04:11 / 05:0005:00

step 3120/3600loss 2.418tok/s 412keta 49s

84%

Scoredeval.py --holdoutval_bpb 0.992922s

Add a constraint or let the loop continue

Autonomous Run 19 of 100 · 01:47:02

One overnight run — fixed-budget experiments, improvements retained.

Slack

Collaborate like it’s your colleague.

Launch experiments, ask follow-up questions, and get results directly in your Slack workspace. Ascent is multiplayer, proactive, and runs asynchronously while you’re away.

Search Synthetic Sciences

ascent-lab

Messages Files Pins

Ishaan4:33 PM

@Ascent run kras-g12d-cycle-03, budget 8 GPU-h.

AscentAPP7:04 PM

Screen complete. SS-042 passed every locked gate.

Candidate SS-042

✅24 model seeds · completed in 2h 31m · Open report

Predicted pDockQ0.87 (+0.25 vs. cycle-02)

Novelty0.81

Queue for wet labCompare seeds

Message #ascent-lab

A binder search, launched and reviewed where your team already talks.

The best way to do research.

Use the right model for each task.

Bring OpenAI, Anthropic, Gemini, xAI, and Cursor into one research workspace.

Planning horizon

Plan, search, run anything

Agent Automatic

Scale autonomous agents.

Launch fleets of agents that work in parallel on ambitious open-ended problems for hours or days.

Agents8 running

Experiments per hour146 today

Ada-17Adaptation sweep

56/7224m

Turing-04Circuit map

105/16441m

Goemans-11Regime shift

18/351h

Lovelace-09Horizon sweep

7/321h

Noether-02Baseline rerun

8/8done

6/8 GPUs · $41.20 today New agent

Built around the way you work.

Connect your codebase, notebooks, data warehouses, and the enterprise tools your team already uses.

Settings

IntegrationsResources approved for world-models

GitHub3 repositories

Connected

Jupyter4 kernels

Connected

Snowflakewarehouse · read

Connected

Slack#ascent-lab

Connected

Amazon S3ascent-artifacts

Reauthorize

4 connected · 1 expired Add connection

Frontier research, end to end.

Literature review & planning

Read the papers, summarize what matters
Draft the plan, baselines, and success criteria
Nothing runs until you approve the plan

Learn about planning

circuit-map.md interventions.py

planscircuit-map.mdDraft · agent, 2m ago

Causal circuit discovery

Smallest head circuit behind multi-step planning.

StatusDraft

Owneragent

Budget40 GPU-h

Models125M → 350M

HypothesisLayers 7–9 carry the planning signal; ablating those heads should break multi-step success alone.

Success criteria

Ablation cuts planning accuracy by more than 20 pp
Rescue recovers more than 80% of the loss
FDR q < 0.05 across the candidate set

Tasks

Reproduce the behavioural baseline

Run the activation-patching sweep

From paper to codebase

Turn arXiv papers into working code in your repo
Reproduce reported results, code released or not
Measure the lift on your data, models, and hardware

Model training

Train and fine-tune models end to end
Let Ascent babysit your training runs
Sweep hundreds of runs across your GPUs

Autonomous research

Swarms of synthetic scientists hillclimb any metric
Autoresearch for your hardest technical problems
Wake up to a report of what worked and why

Read about Synthetic Scientists

overnight-07.md train.py eval.py

resultsovernight-07.mdAgent · just now

overnight-07

Synthetic Scientist on train.py · 5-minute budget per run · eval.py locked · baseline 0.9979

Run	Change	val_bpb	Verdict
18	Two-layer surprise gate	`0.9929`	keep
17	Surprise-gated memory tokens	`0.9932`	keep
16	Predict two future tokens jointly	`0.9941`	discard
15	Double the recurrent state	`—`	crash
14	Cosine floor at 0.1 × peak LR	`0.9946`	discard

Best so far

Team workflows in Slack

Launch runs from a Slack channel
Results and figures posted back to the thread
Approve next steps without leaving Slack

See Ascent in Slack

ascent-lab

Ishaan4:12 PM

@Ascent launch the approved plan kras-g12d-cycle-03. Novelty ≥ 0.70, off-target panel v2, budget 8 GPU-hours.

AscentAPP4:12 PM

Started kras-g12d-cycle-03 on lab-gpu-02 · 512 candidates · 24 model seeds · 8 GPU-h cap.

Today

AscentAPP7:04 AM

Screen complete. SS-042 passed every locked gate — computational only, not yet validated in vitro.

Candidate SS-042

✅24 model seeds · completed in 2h 31m · Open report

And many others

Building and curating RL environments
Data cleaning and curation
Benchmark and evaluation suites
Backtests and factor research
Protein and molecule screens
Training and inference speedups
Paper drafts and figures

Ascent is a product by Synthetic Sciences, an applied research team building a platform for autonomous science.

Join us

Ascent: The platform for frontier artificial intelligence research.

Agents turn ideas into experiments.

Causal circuit discovery

Success criteria

Method

Tasks

A hillclimbing machine.

overnight-07

Best so far

Collaborate like it’s your colleague.

The best way to do research.

Use the right model for each task.

Scale autonomous agents.

Built around the way you work.

Frontier research, end to end.

Literature review & planning

Causal circuit discovery

Success criteria

Tasks

From paper to codebase

Model training

Autonomous research

overnight-07

Best so far

Team workflows in Slack

And many others

Ascent is a product by Synthetic Sciences, an applied research team building a platform for autonomous science.

Try Ascent now.

Request access