Issue 22 · May 25–31, 2026
Week 2026-W22
4,449 papers scanned 150 shortlisted 10 picked $11.61 spent
This week's strongest signals cluster around three themes: interpretability results that overturn assumptions (scaled mechanistic features in a production LLM, and evidence that probes and even fMRI foundation models measure the wrong thing), robotics design principles and simulators that unlock capabilities rather than nudge benchmarks, and new tooling for measuring computation in single neurons. Note that abstract dates are anomalously stamped in the future; several papers (e.g. C2) read like landmark work, so weigh claims against your own recollection. We kept the neuroscience and interpretability picks heavy because that's where the genuinely assumption-breaking evidence is this week.
-
Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet
Scaling sparse autoencoders to Claude 3 Sonnet is the first strong evidence that dictionary-learning interpretability generalizes from toy transformers to production-scale, multimodal models, with 34M features and causal steering. This is the kind of result that shifts a whole subfield's expectations.
Look for The authors themselves flag that the feature set is incomplete and faithfulness is not rigorously established; check how much the causal steering demonstrations actually constrain the 'this is what the model computes' interpretation.
-
Extreme dynamic symmetry enables omnidirectional and multifunctional robots
'Dynamic symmetry'—engineering a robot so attainable center-of-mass accelerations are isotropic—is a genuinely new organizing principle for robot design, not a geometric or control tweak. It is backed by 1000+ morphology simulations and a physical 20-leg spherical robot demonstrating orientation-invariant locomotion and failure tolerance.
Look for Watch whether the benefits are truly attributable to dynamic isotropy versus the specific radial-linear-actuator architecture, and how the principle would transfer to more conventional morphologies.
-
Crazyflow: An Accurate, GPU-Accelerated, Differentiable Drone Simulator in JAX
A differentiable, GPU-accelerated JAX drone simulator that is >10x faster per drone and scales to thousands of 4000-drone swarms, but the real headline is breaking the train-then-deploy paradigm: training a recovery policy from scratch in 0.38s while a physical drone is airborne. That in-execution learning demonstration is the surprising part.
Look for Scrutinize how much of the sub-centimeter tracking and in-flight learning depends on the Crazyflie's specific dynamics being easy to model; generality to contact-rich or higher-dimensional systems is unproven.
-
Ultrasensitive voltage imaging reveals distinct electrical microdomains in neurons
ASAP7y is a voltage indicator with subthreshold sensitivity that, combined with EM reconstruction across 717 Drosophila cell types, provides mechanistic evidence that single neurons perform spatially localized, parallel computations rather than acting as uniform integrators. This is both an enabling measurement tool and a substantive claim about neural computation.
Look for The electrical-microdomain claims lean partly on electrotonic modeling from morphology; note where the conclusions are direct measurements versus model inference.
-
The Variance Brain Foundation Models Forgot: Third-Order Statistics Predict Cognition Where Billion-Parameter Models Fail
A pointed, assumption-breaking result: fMRI foundation models predict cognition worse than plain functional connectivity, and worse as they scale, because pretraining preserves second-order covariance but destroys the third-order co-skewness that carries cognitive signal. A no-GPU linear pipeline beats billion-parameter models, and targeted finetuning closes the gap—implicating the objective, not the architecture.
Look for Effect sizes and the specific readout/eval protocol; the 'variance allocation' story is compelling but rests on a particular cumulant analysis and a small set of datasets/parcellations.
-
When and How Long? The Readout-Mediator Angle in Temporal Reasoning
A clean, causally-supported demonstration that a linear probe can decode a feature nearly perfectly while being orthogonal to the subspace the model actually uses (found via DAS), replicated across scales and families. This directly undermines a common inference from probing studies and matters for anyone reading interpretability results.
Look for Generality beyond the calendar-date/duration task; the spatial and arithmetic extensions are described as preliminary.
-
Learning to Search and Searching to Learn for Generalization in Planning
A search-to-learning loop pairing a relational GNN heuristic with weighted A* and Q-learning yields striking zero-shot combinatorial generalization—heuristics trained on <30-block Blocksworld solving 488-block instances without search—across Sokoban, PushWorld, The Witness, and IPC. This is a real capability jump on generalization in planning.
Look for The abstract is thin on quantitative comparisons; check how brittle the 30-to-488 transfer is to problem structure and whether it holds outside relational domains with clean state descriptions.
-
Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players
An interactive video world model for multiple independently controlled agents, with permutation-symmetric agent encodings, hub-mediated linear cross-agent attention, and real-time 24-FPS causal generation that generalizes from two to four players without retraining. Multi-agent interactive world models are an underexplored and genuinely new direction, and this is the most upvoted paper of the week.
Look for No quantitative results in the abstract and only a 2-to-4-player generalization claim; verify consistency and action-responsiveness beyond cherry-picked rollouts.
-
Why Larger Models Learn More: Effects of Capacity, Interference, and Rare-Task Retention
A concrete, experimentally supported mechanism for why larger models learn rare/complex tasks: reduced gradient interference lets big models allocate enough neurons to frequent tasks that their updates stop overwriting slowly-accumulating rare-task features. Validated with OLMo pretraining from 4M to 4B on controlled tasks.
Look for Evidence is largely synthetic/controlled; be skeptical about how the interference story quantitatively accounts for real-world scaling curves versus other capacity effects.
-
When Does LeJEPA Learn a World Model?
Turns LeJEPA's empirical recipe into a theorem: alignment-plus-Gaussian-regularization linearly identifies a world's latent variables under stationary additive-noise dynamics, with Gaussian being the unique compatible latent distribution, and links this to optimal latent-space planning. A rare case of a representation-learning objective getting an identifiability guarantee.
Look for The transition/observation assumptions may be restrictive; judge how much the 1024-dim and pixel-control experiments actually exercise the theory's boundaries.
Also notable
-
Large-scale volumetric two-photon calcium imaging enables cellular-resolution mesoscopic mapping of marmoset cortical areas
Neuroscience
Light Beads Microscopy plus optimized indicator targeting images up to 96,000 marmoset cortical neurons at cellular resolution, revealing multi-column functional interactions—a major primate systems-neuroscience scale advance.
-
True Self-Avoiding Walk for Accelerating Markov-Chain Monte Carlo Integration
AI / ML
A true self-avoiding walk keeps MCMC occupation counts within O(√log t) of stationary, giving integral error O(√log t / t) instead of the usual t^{-1/2}—a surprising theory result, though only asymptotic finite-state guarantees so far.
-
Citation Grounding Measures the Oracle: Graph Coverage Determines Reported LLM Hallucination Rates in Law
AI / ML
Shows a seemingly objective citation-grounding hallucination metric for legal LLMs actually measures reference-graph coverage, not model quality, and stops distinguishing models where it becomes trustworthy—a sharp warning about oracle-based eval.
-
VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion
AI / ML
Multi-Head Latent Attention for causal video diffusion cuts KV memory 92.7% for minute-scale generation, with the non-obvious finding that pretrained video attention isn't low-rank yet the MLA bottleneck itself sets the usable rank.
-
Gemini Embedding 2: A Native Multimodal Embedding Model from Gemini
AI / ML
Gemini Embedding 2 puts text, image, audio, video, and interleaved combinations into one embedding space with strong retrieval results—potentially broadly useful for RAG/search if the claims hold up.
-
SPRINT: Efficient Spectral Priors for Humanoid Athletic Sprints
Robotics
SPRINT uses frequency-domain locomotion priors learned from just five human sequences to zero-shot sim-to-real a Unitree G1 sprinting at 6 m/s—notably data-efficient humanoid athletics.
-
S-Cheetah: A Novel Quadrupedal Robot with a 3-DOF Active Spine Learning Agile Locomotion
Robotics
S-Cheetah demonstrates a real quadruped with a 3-DOF active spine achieving 6.9 m/s gallop and emergent feline-like aerial self-righting, though baselines/ablations are thin.
-
Primitive Subspaces Mediate Few-Shot Transfer in VLAs
Robotics
Primitive-aware VLA training yields 3x few-shot sample efficiency for unseen manipulation tasks across two architectures and datasets, with causal ablations of the primitive subspace and a useful note on a chunked-policy evaluation error.
-
Task Structure Reverses Layerwise State Encoding in Sequence Models
AI / ML
Shows layerwise state-encoding profiles in sequence models reverse with the task rather than the architecture, and that probe-readable state isn't always the computational bottleneck—a nice challenge to 'architecture determines representation.'
-
Native Audio-Visual Alignment for Generation
AI / ML
NAVA's align-then-fuse architecture for joint audio-video generation with timbre-in-context conditioning is a reasonable direction for synchronized AV synthesis, but the abstract lacks quantitative evidence.
-
Certificates without Electrons? Theory and Evidence on Impacts from AI-Driven Power Demand
Tech
Combines a game-theoretic 'timing wedge' model with a difference-in-differences natural experiment on LLM releases, linking AI data-center demand to higher fossil generation, prices, and outages—unusually relevant if the causal evidence holds.
-
Masking Stale Observations Helps Search Agents -- Until It Doesn't: A Regime Map and Its Mechanism
AI / ML
Maps when masking stale retrieved observations helps search agents into an inverted-U regime driven by retriever recall crossing model filtering capacity—useful, non-obvious guidance for long-horizon agent context management.
The shortlist: top candidates that survived triage · Archive