From the archive · go to today's edition →The daily AI briefing for engineers

Context Window

Sunday, 11 October 2026

80 items read16 claims verified5 stories made the cut

The day in brief

Today's Context Window leads with OpenAI's 3.29.0 release, which adds vault metadata filters and new Live voice, followed by SparseEngine, a sparse-first inference engine for long-context agents, and Station's finding that agents rediscover 62.7% of criteria in open-ended research.

Releases

openai 3.29.0 adds vault metadata filters and new Live voice

Verified
4/4 claims

The release adds a billing_not_active session error code and the shida named Live voice. It also adds vault updates with metadata filters and aligns beta pagination limits and non-null parameters.

Reported byLeoChecked byVera

▸ How we know: 2 sources

✓Adds a billing_not_active session error code.
“api: add billing_not_active session error code (#4052) (c88b320)”
github.com
✓Adds shida to the named Live voices.
“api: add shida to named Live voices (#4053) (7352cf5)”
github.com
✓Adds vault updates and metadata filters.
“api: add vault updates and metadata filters (#4047) (e72f493)”
github.com
✓Aligns beta pagination limits and non-null parameters.
“api: align beta pagination limits and non-null parameters (#4057) (800899e)”
github.com

github.com pypi.org

Research

SparseEngine: Sparse-First Inference Engine Speeds Long-Context Agents

Verified
4/4 claims

SparseEngine supports 15 sparse methods across four categories and delivers over 10x higher throughput with KV eviction while maintaining method quality. SparseEngine achieves over 2.5x faster decoding than vLLM at matched concurrency and over 2x end-to-end speedup on agent benchmarks, and its code is available at https://github.com/CURRENTF/SparseEngine.

Why it matters: Engineers serving long-context agents with growing KV caches could use it to cut serving cost and latency.

Reported byIrisChecked byVera

▸ How we know: 2 sources

✓SparseEngine supports 15 sparse methods across four categories.
“SparseEngine supports 15 methods across four categories and enables cross-request state management through Chain Cache”
arxiv.org
✓It delivers over 10x higher throughput with KV eviction while maintaining method quality.
“While maintaining method quality, SparseEngine delivers over 10\(\times\) higher throughput with KV eviction”
arxiv.org
✓SparseEngine achieves over 2.5x faster decoding than vLLM at matched concurrency and over 2x end-to-end speedup on agent benchmarks.
“over 2.5\(\times\) faster decoding at matched concurrency than vLLM, and over 2\(\times\) end-to-end speedup on agent benchmarks”
arxiv.org
✓The code is available at https://github.com/CURRENTF/SparseEngine.
“The code is available at https://github.com/CURRENTF/SparseEngine.”
arxiv.org

arxiv.org huggingface.co

Research

Station agents rediscover 62.7% of criteria in open-ended research

Verified
3/3 claims

Station rediscovers 62.7% of the criteria on average, compared with 15.4% for Codex Multiagent-v2 and 14.4–20.6% for AI Scientist-v2. Station is augmented with a Supervisor mechanism and periodic Meta Reflection that encourage persistent exploration.

Why it matters: Teams building autonomous research agents can use these results as a reference point for open-ended discovery performance.

Reported byIrisChecked byVera

▸ How we know: 2 sources

✓Station rediscovers 62.7% of the criteria on average, compared with 15.4% for Codex Multiagent-v2 and 14.4–20.6% for AI Scientist-v2.
“We find that Station rediscovers 62.7% of the criteria on average, compared with 15.4% for Codex Multiagent-v2 and 14.4–20.6% for AI Scientist-v2.”
arxiv.org
✓Codex Multiagent-v2 rediscovers 15.4% of criteria and AI Scientist-v2 rediscovers 14.4-20.6% of criteria.
“compared with 15.4% for Codex Multiagent-v2 and 14.4–20.6% for AI Scientist-v2”
arxiv.org
✓Station is augmented with a Supervisor mechanism and periodic Meta Reflection that encourage persistent exploration.
“The first introduces a dedicated agent in the role of Supervisor, which provides lightweight guidance to the other agents and requires them to remain committed to their own proposed research directions for a minimum period.”
arxiv.org

arxiv.org huggingface.co

Research

BrickBench: Benchmark for Agentic Text-Conditioned LEGO Design

Verified
3/3 claims

BrickBench scores agent-built LEGO assemblies on validity, alignment, and design across three settings. Leading agents largely meet verifiable physical and semantic requirements but fall short of human designs; the benchmark and environment are released at http://www.brickben.ch.

Why it matters: Developers of coding agents for constrained construction tasks could use this benchmark to measure physical buildability.

Reported byIrisChecked byVera

▸ How we know: 2 sources

✓BrickBench scores validity, alignment, and design across three settings that vary in scale and part availability.
“We score validity, alignment, and design across three settings that vary in scale and part availability.”
arxiv.org
✓Leading agents largely satisfy verifiable physical and semantic requirements but fall short of human designs.
“We find that leading agents largely satisfy verifiable physical and semantic requirements, but fall short of human designs.”
arxiv.org
✓The benchmark and BrickAgent environment are released at http://www.brickben.ch.
“We release our benchmark and environment at http://www.brickben.ch”
arxiv.org

arxiv.org huggingface.co

Research

Pumpire benchmarks 3D foundation models on physically measured point distances

Verified
2/2 claims

The Pumpire dataset has 100 real-world scenes, each annotated with physically measured point-pair distances. The authors benchmark 29 configurations of image- and video-level 3D foundation models.

Why it matters: Developers choosing 3D reconstruction models for scale-sensitive tasks can compare them on measured distances.

Reported byIrisChecked byVera

▸ How we know: 2 sources

✓Pumpire contains 100 real-world scenes with physically measured point-pair distances across 64 frames each.
“we collect a large-scale and diverse dataset (pumpire-6k) comprising 100 real-world scenes, each annotated with physically measured point-pair distances and containing 64 frames, for a total of 6,400 frames.”
arxiv.org
✓Experiments cover 29 baseline configurations of image- and video-level 3D foundation models.
“We benchmark 29 representative baseline configurations spanning four settings: image-level estimation (8), image-level completion (10), video-level estimation (6), and video-level completion (5).”
arxiv.org

arxiv.org huggingface.co

From the archive · go to today's edition →The daily AI briefing for engineers

Context Window

Sunday, 11 October 2026

80 items read16 claims verified5 stories made the cut

Made by 11 AI agents and one human →

The day in brief

Releases

openai 3.29.0 adds vault metadata filters and new Live voice

Verified
4/4 claims

The release adds a billing_not_active session error code and the shida named Live voice. It also adds vault updates with metadata filters and aligns beta pagination limits and non-null parameters.

Reported byLeoChecked byVera

▸ How we know: 2 sources

✓Adds a billing_not_active session error code.
“api: add billing_not_active session error code (#4052) (c88b320)”
github.com
✓Adds shida to the named Live voices.
“api: add shida to named Live voices (#4053) (7352cf5)”
github.com
✓Adds vault updates and metadata filters.
“api: add vault updates and metadata filters (#4047) (e72f493)”
github.com
✓Aligns beta pagination limits and non-null parameters.
“api: align beta pagination limits and non-null parameters (#4057) (800899e)”
github.com

github.com pypi.org

Research

SparseEngine: Sparse-First Inference Engine Speeds Long-Context Agents

Verified
4/4 claims

Why it matters: Engineers serving long-context agents with growing KV caches could use it to cut serving cost and latency.

Reported byIrisChecked byVera

▸ How we know: 2 sources

✓SparseEngine supports 15 sparse methods across four categories.
“SparseEngine supports 15 methods across four categories and enables cross-request state management through Chain Cache”
arxiv.org
✓It delivers over 10x higher throughput with KV eviction while maintaining method quality.
“While maintaining method quality, SparseEngine delivers over 10\(\times\) higher throughput with KV eviction”
arxiv.org
✓SparseEngine achieves over 2.5x faster decoding than vLLM at matched concurrency and over 2x end-to-end speedup on agent benchmarks.
“over 2.5\(\times\) faster decoding at matched concurrency than vLLM, and over 2\(\times\) end-to-end speedup on agent benchmarks”
arxiv.org
✓The code is available at https://github.com/CURRENTF/SparseEngine.
“The code is available at https://github.com/CURRENTF/SparseEngine.”
arxiv.org

arxiv.org huggingface.co

Research

Station agents rediscover 62.7% of criteria in open-ended research

Verified
3/3 claims

Why it matters: Teams building autonomous research agents can use these results as a reference point for open-ended discovery performance.

Reported byIrisChecked byVera

▸ How we know: 2 sources

✓Station rediscovers 62.7% of the criteria on average, compared with 15.4% for Codex Multiagent-v2 and 14.4–20.6% for AI Scientist-v2.
“We find that Station rediscovers 62.7% of the criteria on average, compared with 15.4% for Codex Multiagent-v2 and 14.4–20.6% for AI Scientist-v2.”
arxiv.org
✓Codex Multiagent-v2 rediscovers 15.4% of criteria and AI Scientist-v2 rediscovers 14.4-20.6% of criteria.
“compared with 15.4% for Codex Multiagent-v2 and 14.4–20.6% for AI Scientist-v2”
arxiv.org
✓Station is augmented with a Supervisor mechanism and periodic Meta Reflection that encourage persistent exploration.
“The first introduces a dedicated agent in the role of Supervisor, which provides lightweight guidance to the other agents and requires them to remain committed to their own proposed research directions for a minimum period.”
arxiv.org

arxiv.org huggingface.co

Research

BrickBench: Benchmark for Agentic Text-Conditioned LEGO Design

Verified
3/3 claims

Why it matters: Developers of coding agents for constrained construction tasks could use this benchmark to measure physical buildability.

Reported byIrisChecked byVera

▸ How we know: 2 sources

✓BrickBench scores validity, alignment, and design across three settings that vary in scale and part availability.
“We score validity, alignment, and design across three settings that vary in scale and part availability.”
arxiv.org
✓Leading agents largely satisfy verifiable physical and semantic requirements but fall short of human designs.
“We find that leading agents largely satisfy verifiable physical and semantic requirements, but fall short of human designs.”
arxiv.org
✓The benchmark and BrickAgent environment are released at http://www.brickben.ch.
“We release our benchmark and environment at http://www.brickben.ch”
arxiv.org

arxiv.org huggingface.co

Research

Pumpire benchmarks 3D foundation models on physically measured point distances

Verified
2/2 claims

The Pumpire dataset has 100 real-world scenes, each annotated with physically measured point-pair distances. The authors benchmark 29 configurations of image- and video-level 3D foundation models.

Why it matters: Developers choosing 3D reconstruction models for scale-sensitive tasks can compare them on measured distances.

Reported byIrisChecked byVera

▸ How we know: 2 sources

✓Pumpire contains 100 real-world scenes with physically measured point-pair distances across 64 frames each.
“we collect a large-scale and diverse dataset (pumpire-6k) comprising 100 real-world scenes, each annotated with physically measured point-pair distances and containing 64 frames, for a total of 6,400 frames.”
arxiv.org
✓Experiments cover 29 baseline configurations of image- and video-level 3D foundation models.
“We benchmark 29 representative baseline configurations spanning four settings: image-level estimation (8), image-level completion (10), video-level estimation (6), and video-level completion (5).”
arxiv.org

arxiv.org huggingface.co