AI Paper Digest · August 26, 2026
AI Paper Digest · August 26, 2026
Weekly selected AI frontier papers, insight into academic cutting-edge.
1. Recuris: Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses
Institutions: Gen-Verse, multi-institutional collaboration (arXiv 2026)
Problem: Recursive self-improvement (RSI) remains difficult in long-horizon tasks, where growing interaction histories obscure task state and misalign skill invocation with current needs. Existing approaches struggle to let agents continuously transform accumulated experience into increasingly effective behavior.
Method: The paper introduces Recuris, a recursive Experiential-Working Memory architecture. Working Memory tracks task progress and guides skill selection from Experiential Memory, grounding skill use in current needs rather than the full history. Execution itself generates structured evidence that localizes failures to specific memory components. A fixed Meta-Agent then turns that evidence into localized, validation-gated updates to Skill Memory, forming a bounded recursive memory-evolution loop.
Key Finding: Across four long-horizon benchmarks and ten models, Recuris improves task success in 35 of 37 completed model-benchmark pairs, carrying frontier models to SOTA-level performance: on tau-bench it adds +17.8 points to GPT-5.6 Sol and +15.6 to Claude Opus 5 (reaching 87.9%), and +16.6/+13.5 points on Qwen3.6-27B/35B on SkillFlow. The advantage widens with interaction horizon — +32.2 points on the longest tasks — and common long-horizon failures drop by up to 80%.
Link: http://arxiv.org/abs/2608.24876
2. SPO++: Stream-Aligned Policy Optimization for Asynchronous Agentic RL
Institutions: Machine learning and reinforcement learning research community (arXiv 2026)
Problem: Group-relative reinforcement learning methods (e.g., GRPO) must wait for sibling rollouts of the same prompt before computing advantages, which is costly for long and variable tool-use trajectories. While Single-stream Policy Optimization (SPO) removes this dependency via persistent prompt-level value estimates, its advantage normalization creates a mathematical mismatch.
Method: The paper proposes SPO++, which introduces two key fixes over SPO: (1) it proves that trajectory-level centering generally does not center the token-weighted quantity consumed by the actor, and fixes this by standardizing terminal-outcome advantages under the action-token measure; (2) it organizes prompt evidence by the policy event that generated it rather than the learner’s receipt order.
Key Finding: Across matched runs on ALFWorld at two model scales and on Math-TIR, SPO++ improves online learning efficiency over SPO. A paired ablation identifies action-token-measure normalization as the strongest tested component, establishing a more efficient training paradigm for asynchronous agentic RL.
Link: http://arxiv.org/abs/2608.24870
3. LAION-BVD: A 10-Million-Hour Open Video Dataset for Multimodal Pre-training
Institutions: LAION research team, multi-institutional collaboration (arXiv 2026)
Problem: Multimodal video pre-training has long been constrained by data scale — open-source video datasets typically contain only tens of thousands of hours, while massive closed datasets (like YouTube-15M) are not publicly available. The field needs a truly large-scale, openly accessible video dataset to advance video understanding and audio-visual learning.
Method: The researchers collected 1.3B platform-specific video URLs from CommonCrawl, successfully downloading 80M videos totaling 10 million hours of content. Using content-aware scene detection, they extract key clips and synthetically generate video and audio captions. They also explore video frames as an alternative source of image-text data by extracting scene-changing frames, which exhibit a visual distribution distinct from standard web image corpora.
Key Finding: Models trained on LAION-BVD achieve competitive performance on video-text and audio-text benchmarks, with consistent improvements as training or model scale increases. The dataset’s video frames also achieve strong image-text retrieval performance. LAION-BVD is released to the research community, significantly expanding open access to multimodal video data at an unprecedented scale.
Link: http://arxiv.org/abs/2608.24845
Compiled from multiple academic sources.
