Research/OpenAI News/July 29, 2026 at 3:00 PM

How enabling two settings tripled our scores on the ARC-AGI-3 benchmark

How two API settings improved GPT-5.6 performance on ARC-AGI-3, boosting scores and efficiency by retaining reasoning and enabling compaction.

ResearchAI

Research / OpenAI News

Source

Follow OpenAI News to make it a durable For You signal.

Fast Brief

OpenAI News / Research

1 min read

1How two API settings improved GPT-5.6 performance on ARC-AGI-3, boosting scores and efficiency by retaining reasoning and enabling compaction.

research

Source Lens

High-credibility vendor_blog source with strong edition heat and sparse entity coverage.

High confidence

Source: OpenAI News
Type: vendor_blog
Credibility: 95/100
Heat: 84/100
Signals: 0 entities / 1 tag

Article Corroboration

2 independent stories from 2 sources corroborate this article around research.

Corroborated

Sources: 3
Matches: 2
Signals: 2
Avg trust: 86

arXiv paper: TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM1 signal / 86 trust

arXiv AI/ML / Researchresearch

Preview actions train this device only. Live stories will sync once production news IDs are available.

ReadSource

Microsoft AI bets on cheap specialist models instead of chasing the frontier1 signal / 78 trust

The Decoder / Researchresearch

Preview actions train this device only. Live stories will sync once production news IDs are available.

ReadSource

Server Profile

Server-side profile learning will appear after saved stories, reads, source clicks, or Less feedback sync.

Server Waiting

Trained: 0
Ignored: 0
Hidden: 0

Reader Fit

No saved reader signals match this article yet; 3 follow-ups keep the thread moving.

Discovery Read

Profile matches: 0
Follow-ups: 3
Reader bias: Balanced

Discovery

This article is training a new reader profile.

Broaden Topic / 2 signals / 108 scoreMicrosoft AI bets on cheap specialist models instead of chasing the frontierSame topic

Preview actions train this device only. Live stories will sync once production news IDs are available.

ReadSource

Read Training

Article read training starts after the reader opens this story.

Awaiting Read

Opened: No
Training signals: 0
Next target: Open

OpenedNext / Open
The article open is logged, but it does not train preferences yet.
Meaningful readWait / 35%
Reading to 35% starts profile training and reading history.
Deep readWait / 80%
Reading to 80% strengthens related topic and entity memory.

Open this article to start reader-memory training for Research.

Feedback Loop

Article feedback loop will appear after an explicit action.

Waiting

Action: None
Signal delta: 0
Bias shift: 0
Topic: Research

Awaiting feedback

Deep read, save, share, or press Less to show article-level training feedback.

Learning Impact

This article can start reader-memory signals and stage 3 follow-up recommendations.

Learning Ready

Article memory: 0
Save adds: +2
Source adds: +1
Less removes: 0
Next candidates: 3

Read+1 signal / +0.3 bias

Read would start a new Research memory; angle memory waits for a deeper read.

Save+2 signals / +0.6 bias

Save would add Research, OpenAI News to the reader profile.

Share+2 signals / +0.9 bias

Share would add Research, OpenAI News and push freshness and novelty harder.

Source+1 signal / +0.6 bias

Source would add OpenAI News to the reader profile.

Less0 signals / -0.4 bias

Less would only dampen ranking bias until a signal exists.

Next Recommendation

Same topic / 2 signals / 108 scoreMicrosoft AI bets on cheap specialist models instead of chasing the frontier

Preview actions train this device only. Live stories will sync once production news IDs are available.

ReadSource

Same topic / 2 signals / 106 scorearXiv paper: TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM

Preview actions train this device only. Live stories will sync once production news IDs are available.

ReadSource

Next Reads

4 next reads: 4 continue, 0 explore, and 0 verify.

Next Reads Ready

Candidates: 4
Continue: 4
Explore: 0
Verify: 0

Microsoft AI bets on cheap specialist models instead of chasing the frontierContinue

Research / The Decoder / 2 signals / 108 scoreSame topic

Preview actions train this device only. Live stories will sync once production news IDs are available.

ReadSource

arXiv paper: TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAMContinue

Research / arXiv AI/ML / 2 signals / 106 scoreSame topic

Preview actions train this device only. Live stories will sync once production news IDs are available.

ReadSource

arXiv paper: Mental World ModelingContinue

Research / arXiv AI/ML / 2 signals / 106 scoreSame topic

Preview actions train this device only. Live stories will sync once production news IDs are available.

ReadSource

arXiv paper: From Classification to Regression: Using a Fruitfly to Solve EquationsContinue

Research / arXiv AI/ML / 2 signals / 106 scoreSame topic

Preview actions train this device only. Live stories will sync once production news IDs are available.

ReadSource

Article Signals

Topic: Research
Source: OpenAI News
Entities: None
Tags: research
Reader bias: F1.0 / N1.0

Reading Path

5 follow-ups ranked by article overlap, shared entities, and reader signals.

2 signals / 108 scoreMicrosoft AI bets on cheap specialist models instead of chasing the frontierSame topic

Preview actions train this device only. Live stories will sync once production news IDs are available.

ReadSource

2 signals / 106 scorearXiv paper: TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAMSame topic

Preview actions train this device only. Live stories will sync once production news IDs are available.

ReadSource

2 signals / 106 scorearXiv paper: Mental World ModelingSame topic

Preview actions train this device only. Live stories will sync once production news IDs are available.

ReadSource

2 signals / 106 scorearXiv paper: From Classification to Regression: Using a Fruitfly to Solve EquationsSame topic

Preview actions train this device only. Live stories will sync once production news IDs are available.

ReadSource

2 signals / 106 scorearXiv paper: VidMap: Exploiting Temporal Structure for Video-Based Structure-from-MotionSame topic

Preview actions train this device only. Live stories will sync once production news IDs are available.

ReadSource