Research Briefing — June 26, 2026

ABC is a fully open-source behavior-cloning stack centered on ABC-130K, the largest open teleoperation dataset to date with 3,500 hours over 130K episodes across 195 tasks, plus open hardware, training infrastructure, simulation, 400 hours of sim-teleop data, and a co-training recipe correlating sim and real evaluation. The authors compare DiT and VLA architectural choices. The scale and full openness make this a potentially field-shaping resource.

This work shows that hallucination in generative world models concentrates in low-coverage regions of state-action space and is therefore predictable and preventable with lightweight data-centric signals. It introduces MMBench2, a 427-hour 210-task dataset, trains a 350M world model, and identifies three distinct hallucination modes with targeted mitigations.

This paper proves that gradient equilibrium (a recently introduced online optimization framework generalizing first-order stationarity) is algorithmically equivalent to Blackwell approachability. Any approachability problem can be solved via black-box GEQ oracle queries with no asymptotic loss, and vice versa.

RiVER trains LLMs on score-based optimization tasks without ground-truth solutions, using deterministic execution feedback as continuous-valued reward instead of verifiable answers. It addresses scale dominance and frequency dominance challenges that arise when applying group-relative RL to such continuous rewards.

This theoretical work reframes neural networks as models of computation rather than only flexible basis functions, providing complexity-theoretic approximation rates and a characterization of universal approximation tied to circuit computability. It offers explicit depth/width/parameter bounds for functions computable by real-valued circuits.

Autoregressive Boltzmann Generators introduce an autoregressive modeling framework for sampling molecular systems at thermodynamic equilibrium, overcoming the expressivity and likelihood-cost limitations of normalizing-flow-based Boltzmann Generators. It enables rapid generation of uncorrelated equilibrium samples with exact likelihoods.

EpiKV is a training-free KV cache eviction method that scores tokens by the change in the model's internal representation (the epiphany score) rather than attention weights, avoiding materialization of the attention matrix. This makes it compatible with fused FlashAttention kernels in production for long reasoning traces.

This work introduces LCS-Bench, a theory-scale auto-formalization benchmark covering 327 interdependent textbook items in Logics for Computer Science, built via a semi-automated agentic pipeline with concept graphs, signature planning, and counterexample-driven sorry-filling. It targets the open challenge of coherently formalizing entire theories rather than isolated statements.

This paper develops the first generalization theory for JEPA-based world models, formulating JEPA pretraining as conditional spectral graph learning equivalent to low-rank factorization of an action-conditioned co-occurrence matrix. It connects pretraining error to downstream planning regret.

This large-scale study characterizes AI nudification on 4chan, identifying over 24,000 non-consensual sexually explicit synthetic images and revealing that non-celebrity targets now account for 55.8 percent versus 4.7 percent in prior studies. It documents the technology and community dynamics enabling this harm.

This work introduces variable-delay real-time RL where the environment continues progressing while the agent deliberates, making the cost of planning explicit. They train a lightweight gating policy on top of a planner to choose state-dependent planning budgets, avoiding the paralysis of naively planning how long to plan.

This work shows that high-quality reasoning SFT examples can be identified using only the first 100 reasoning tokens, via the loss at a randomly perturbed pretrained checkpoint, avoiding costly filtering with strong reasoning models. It demonstrates that difficult problems are detectable early, enabling cheaper data curation for reasoning models.