Also Worth Noting - 2026-08-17

Also Worth Noting

02 [Open-source] DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data Competitive English benchmark performance at 1B parameters does not require non-permissible corpora. Mimir v1 uses the Hierarchical Reasoning Model architecture trained from scratch on a mixture of 161 permissible datasets, outperforming the original HRM-Text 1B and setting a new state of the art for Danish. The direct implication: teams constrained by licensing can now point to a concrete architecture-and-data recipe rather than treating permissible-only training as a performance ceiling. link

03 [Agent] Second Thought: Reasoning in Parallel as LLM Agents Act and Observe ReAct agents freeze their reasoning the moment an action fires, leaving the action-observation wait interval completely idle. Second Thought forks four auxiliary reasoning branches the instant each Thought phase ends and decodes them concurrently with the environment call, feeding useful conclusions into the next turn. No base model changes, no fine-tuning. Any team running a ReAct pipeline today can treat this as a drop-in throughput upgrade for reasoning quality without touching the underlying model. link

04 [Eval] Claim-Level Reliability Assessment for Efficient Test-Time Reasoning Sampling more solutions is the default test-time scaling move, but it dilutes the signal by averaging over routine tokens rather than isolating the steps that actually determine correctness. CLR condenses each reasoning trace into a compact set of decision-critical claims, then reallocates compute to verifying those claims rather than generating additional full traces. On math benchmarks it matches or beats majority-vote accuracy at lower compute cost, making it a practical alternative for teams already hitting diminishing returns from repeated sampling. link

05 [Training] Dion3: Full-Stack Orthogonal Updates Muon's orthogonalization step runs in cubic time, and weight sharding compounds that cost enough to erase the optimizer's gradient quality advantage in distributed settings. Dion3 attacks this at every level: a Gram Newton-Schulz algorithm cuts the FLOP count, CuteDSL kernels exploit symmetry to accelerate the computation, and a megabatching strategy reduces cross-shard communication. Teams evaluating orthogonal-update optimizers for large distributed runs now have a version where the overhead argument no longer holds. link

06 [Application] Forecast Collapse in Time-Series Foundation Models Across 1,000 US equities, time-series foundation models produce predictions that flatten to near-constant values with poor cross-sectional ranking, yet the same models forecast trading volume without collapsing. The failure tracks target predictability, not a general model defect, which means standard benchmark scores on volume or other predictable series will not warn you before deploying on returns. Teams using TSFMs for financial signals should test cross-sectional correlation on their specific target distribution before treating benchmark numbers as deployment proxies. link