Research today clusters around two practical themes rather than new capability claims. One group of papers chips away at inference cost: faster speculative decoding, parameter-free teacher-student distillation, and a finding that offline student rollouts often beat on-policy distillation outright. A second group asks whether agents can turn raw experience, demonstrations, or competitive play into reusable skill, with Opus5.5 running on Claude Code taking six gold medals in a long-running adversarial game benchmark while still losing most human ladders. A third thread is methodological hygiene: two papers show biological and spatial reasoning models lean on their language priors more than on the structured inputs they are supposed to use, and a third introduces a benchmark for data-video understanding that most multimodal models still fail. On the industry side, Karpathy's comment about a widening gap in shared understanding of LLM capability drew outsized attention, while QGIS's ban on AI-assisted contributions and Silicon Valley's uptake of the White House's "superintelligence" rebrand point to diverging reactions to AI inside institutions.
QGIS, the open-source geospatial project, banned AI-assisted code contributions over provenance and review-burden concerns. Separately, a Karpathy post on the widening gap in public understanding of current LLM capability drew outsized attention (492 points), while coverage shows parts of Silicon Valley adopting the White House's push to rebrand AI as "superintelligence." Smaller stories cover data-center siting NDA friction and skepticism that voice AI has reached its breakout moment.
The QGIS open-source geospatial project adopted an enhancement proposal banning AI-assisted code contributions.
A widely-discussed Karpathy post (492 HN points) argues that public understanding of what current LLMs can and can't do is falling further out of sync with their actual capability.
Some of Silicon Valley's most powerful figures are adopting the White House's push to rebrand AI as "super intelligence," even as most of the public keeps using "AI."
Debate over whether ending NDA practices around data center siting deals could ease local opposition and moratoriums.
Executives argue voice AI still misses context in ways that break the broader pipeline, keeping it short of a breakout consumer moment.
Reports indicate Apple's long-rumored smart home product now has a name, HomeView, and a launch date.
Six papers cut inference and training cost without changing base weights. DCOCO restores token dependencies in parallel diffusion drafters for up to 6.6x speedups; PreSTO and a layer-selection method speed speculative decoding further. Universal Textual Teaching distills teacher-student gaps into a reusable text "Primer" with no parameter updates, a 17-pair study finds offline-rollout distillation often beats on-policy distillation, and SynthSelf brings self-distillation to small models.
DCOCO restores intra-block token dependencies for parallel diffusion drafters by scoring candidate tokens' contextual compatibility, reaching up to 6.6x speedups over vanilla decoding.
UTT distills teacher-student knowledge gaps into a reusable natural-language "Primer" with no parameter updates, raising a student's KernelBench accuracy from 9.4% to 48.6%.
Across 17 teacher-student pairs from 1.5B to 235B parameters, distilling from offline student rollouts (Semi-OPD) beats on-policy distillation in 14 of 17 cases.
PreSTO prefetches batches of future Metropolis-Hastings proposals for power sampling, cutting model calls roughly 3.9x while leaving the sampler's transition kernel unchanged.
Entropy- and alignment-based layer-selection for EAGLE3-style speculative decoding outperforms the fixed-position default, with a clearer account of why simple heuristics sometimes stay competitive.
SynthSelf outsources in-distribution rewriting to a minimally larger same-family aligner, letting small models continually learn new capabilities without forgetting, across 4 acquisition datasets and 14 retention benchmarks.
Papers formalize how agents turn experience into reusable competence. AAArena has agents revise game policies from replays; Opus5.5 on Claude Code wins six golds but tops no remaining human ladder, weaker as rule complexity rises. Work2Skill induces reusable GUI skills from human desktop demos. Encore reads a few robot demos as evidence, not training data, to write manipulation code. SoftwareGym synthesizes 182 runnable environments for computer-use agent training.
AAArena benchmarks AI agents revising game-playing policies from replays across 12 adversarial games against 1,920 archived human programs; Opus5.5 with Claude Code wins 6 gold medals but no configuration tops the remaining 6 human ladders.
Work2Skill records screen, mouse/keyboard, and accessibility-tree data from 350 human desktop demonstrations, then induces reusable skills that separate workflow rules from task-specific detail.
Encore distills a few robot demonstrations into keyframes and trajectories read by a coding agent, which writes and freezes a manipulation policy program, outperforming the strongest prior agentic baseline on LIBERO-PRO (96.3% vs 89.3%).
SoftwareGym synthesizes 182 runnable GUI applications across twelve categories with verified functionality, and training on their trajectories lifts Pass@3 on out-of-domain OSWorld from 41.39% to 56.97%.
Two papers find reasoning models score well without actually using inputs built for them. Perturbation tests on six biological reasoning models find leading genomic/protein foundation models contribute little to accuracy, with the LLM following text over biological evidence in over 97% of conflicts. DataVista, the first data-video-understanding benchmark, finds the best model (Gemini-3.1-Pro) reaches only 70% accuracy, weakest on causal reasoning and narrative structure, well below humans.
Testing six biological reasoning models with input perturbations and evidence conflicts finds Evo2 and ESM3 contribute little to BioReason's final accuracy, which instead follows the text input in 97.9-99.7% of conflicts.
DataVista, the first benchmark for data-video understanding, tests 19 MLLMs on 961 real-world data videos and 6,775 questions; the best model reaches only 70.0% accuracy, weakest on causal reasoning and narrative structure.
MetaKernelBench tests whether a natural-language skill distilled from one GPU-kernel DSL attempt improves a fresh attempt in a different DSL; paired lift ranges from -19% to +29% and regressions occur in every model tested.
Four papers push embodied/spatial learning toward structure over scale. Dex-One2Many abstracts one demo video into scene graphs guiding RL for manipulation, widening its lead to 71% unseen. TouchScale's 500-hour visual-tactile dataset lifts manipulation success from 22.5% to 57.5%. SAVER trains sparse-frame video reasoning using dense views only as training-time references. S-Space finds a linear subspace encoding spatial coordinates, exposing viewpoint transformation as the key bottleneck.
Dex-One2Many abstracts one human demo video into scene graphs that constrain and guide RL exploration, transferring zero-shot to a real multi-fingered hand and widening its lead over baselines to 71% in unseen scenarios.
TouchScale is a 500-hour unified-sensor visual-tactile dataset; pretraining on it raises zero-shot contact IoU on an unseen sensor from 0.134 to 0.383 and lifts real-world manipulation success from 22.5% to 57.5%.
SAVER trains sparse-frame video reasoning using dense video as a training-time-only reference, matching or surpassing dense-frame Qwen3.5 baselines after training on just 1,250 grounding examples.
S-Space identifies a linear subspace in multimodal models encoding object coordinates; perception reads locations into it and causal interventions alter spatial judgments, but models fall short on the viewpoint transformations needed to reason across changing perspectives.