AI Paper Digest · September 2, 2026

Weekly selected AI frontier papers, insight into academic cutting-edge.


1. LayerRecall: A State-Conditioned Memory Router for Long-Horizon Consistency in Video Generation

Institutions: Zhejiang University (Yixuan Ding, Jiahao Kong, Wei Huang, Ruijie Quan, Yi Yang, et al.)

Problem: Autoregressive video diffusion enables scalable long-video generation by producing chunks from a bounded recent context. While recency-based caching preserves local continuity, it evicts the historical cues needed when subjects, objects, scenes, or attributes reappear. Existing memory mechanisms expose models to nonlocal history, but access alone does not guarantee effective use — analysis shows video DiT layers have distinctly different preferences for current, recent, and distant context.

Method: The paper introduces LayerRecall, a current-conditioned, layer-selective memory router that retrieves relevant historical K/V states but injects them only into backbone-specific, memory-sensitive layers while preserving local attention elsewhere. To reduce reliance on scarce high-quality long-horizon videos and explicit memory-allocation labels, it proposes Cross-Horizon Prediction Matching (CHPM), which uses a privileged long-context reference to supervise the bounded-memory router in prediction space. Evaluation spans 100 multi-shot prompts.

Key Finding: LayerRecall achieves the best overall results on MemoBench and MovieBench while matching its backbone on VBench-Long — stronger long-range recovery without sacrificing local continuity. Qualitative analyses reveal memory-guided self-correction, where initially mismatched local attributes return to their historical appearance without resetting ongoing motion or scene structure. Additional analyses show cross-backbone portability and negligible inference overhead.

Link: https://arxiv.org/abs/2608.28460


2. LoopArena: Benchmarking Models as Runtime Controllers for Loop Engineering

Institutions: AMAP-ML, multi-institutional collaboration (arXiv 2026)

Problem: One end-to-end run cannot answer the key question: does final success or failure reflect the loop’s guidance or the coding agent’s own ability? “Loop engineering” — where after each coding round a model decides what the agent should do, verify, or whether to stop — had no benchmark that isolates the controller from the worker.

Method: The paper introduces LoopArena, a benchmark for evaluating how well one model can guide a separate, fixed coding agent through a long-running task. The model under evaluation is the Controller: after each coding round it receives a structured summary of the run and instructs a fixed Worker on what to do or verify next, or whether to stop. Three complementary settings: Type I scores next-step Loop Contract selection through execution-validated questions without running the Worker; Type II executes repeated control over a slice of a full task; Type III evaluates the paired full task from its original state.

Key Finding: On full tasks, the best observed Strict Success Rate is just 24.69% — substantial room for improvement in long-horizon loop control. Yet the cost benefit is significant: the paired reduction in estimated inference cost averages 64.4%, and Type II produces a similar model ordering to full-task evaluation under the main criterion (Spearman ρ=0.9747). Benchmark data and evaluation code are open-sourced.

Link: https://arxiv.org/abs/2608.28281


3. Fast Weight Attention for Continual Learning: Recasting Recurrent State Updates as Online Learning Rules

Institutions: Machine learning and theoretical research community (Yifan Zhang, Steve Ta, Jasper Zhang, Jichen Feng, Shuzhen Li, Yongxin Zhang, et al., including Mengdi Wang, Quanquan Gu, Andrew Chi-Chih Yao)

Problem: Recurrent fast-weight memories and selective state-space models compress an expanding context into a fixed-size recurrent state, making the state transition naturally an online learning rule. But the rule had not been systematically derived under autoregressive “read-after-write” semantics — the common same-step association (φ(k_t), v_t) is causal, yet optimizes a different internal objective.

Method: For the prefix-prediction objective, the paper derives normalized first-order updates for squared-error regression and negative inner-product objectives. The regression family comprises Falcon-1 (a scalar NLMS update), Falcon-2 (its per-column extension), and Falcon-3 (a sliding-window mini-batch update), with Falcon-1A/2A/3A as the corresponding inner-product variants. It provides recurrent, masked-parallel, and chunk-parallel forms, together with numerically stable positive-decay renormalization.

Key Finding: Representative variants remain competitive in language modeling and improve length extrapolation on variable-digit addition. The framework cleanly separates temporal alignment, plasticity, forgetting, and bounded rehearsal in recurrent sequence models — a new lens for understanding recurrent memory as online learning.

Link: https://arxiv.org/abs/2608.27763


Compiled from multiple academic sources.