Also Worth Noting - 2026-08-08

Also Worth Noting

02 [Eval] Toward Skill-Native LLMs: Skill Entropy for Benchmarking and Training Long-Horizon Reasoning A model that aces math and aces scheduling separately can still fail when a single reasoning chain requires both, and existing benchmarks have no way to measure that gap. Skill Entropy quantifies the difficulty of cross-skill switching inside a reasoning trajectory, giving practitioners a concrete metric beyond per-skill accuracy. The framework addresses both evaluation and training, targeting the specific failure mode where inter-skill transitions degrade performance that isolated benchmarks never expose. Teams building or evaluating general-purpose reasoning systems should treat cross-skill degradation as a first-class metric. link

03 [Agent] ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment Treating every step in a search trajectory as equally valuable is the training flaw that makes long-horizon search agents inefficient. ABSeeker traces credit backward from the final answer to identify which retrieval and verification steps actually contributed, converting sparse trajectory-level outcomes into fine-grained step-level signals. This answer-backtracked credit assignment (ABC) framework works within both SFT and RL without requiring step-level human annotation. The result is fewer redundant actions in long-horizon search trajectories, a direct win for teams building retrieval-augmented agents at production scale. link

04 [Training] Ego2Robot: Scalable Robot Data Synthesis from Egocentric Human Data Retargeting egocentric human video for robot training has worked at small scale for per-task policies, but whether it transfers as pretraining signal for vision-language-action models at scale was untested. Ego2Robot builds a scalable pipeline that converts human manipulation videos into robot-format data through action retargeting and rendering, then uses the output for VLA pretraining. The findings show meaningful pretraining benefit, suggesting that the vast existing supply of human egocentric video can substitute for expensive robot-specific data collection. For robotics teams facing data bottlenecks, this is a path toward internet-scale pretraining without robot-hours. link

05 [Inference] PaDoc: Layout-Grounded Parallel Decoding for Document Parsing Autoregressive document parsers pay a decoding cost that grows with total page content, even for regions that are structurally independent of each other. PaDoc uses the predicted layout as a branching structure over a shared visual representation, then decodes independent regions in parallel under a region-sufficiency condition, keeping a single visual prefix instead of repeated visual prefills. Decoding length scales with the number of regions rather than total content volume. For teams running high-volume document pipelines, this is a concrete latency reduction without sacrificing the full-page context that two-stage crop-based parsers lose. link

06 [Open-source] K-EXAONE 2.0 Technical Report At 750B total parameters with only 37B activated per token, K-EXAONE 2.0 is the largest open-weight MoE released by a Korean lab and a direct challenge to the assumption that frontier-scale open weights come exclusively from US or Chinese institutions. LG AI Research built it by upcycling the prior K-EXAONE architecture rather than training from scratch, expanding into a MoE structure with more than three times the predecessor's capacity. Context length extends to 256K tokens and multilingual coverage grows from six to ten languages. The open-weight release at this scale shifts what non-US labs can offer practitioners who need frontier-class models without API dependency. link