System One decision models produce calibrated probabilities for legal authority tasks, enabling abstention when the correct authority is missing. By extending CaseHOLD and comparing multiple decision models with several LLMs, the work shows abstention reduces wrong citations and that distilling probabilities into a smaller model yields effective performance at low cost; emphasis is on reducing incorrect citations rather than pure accuracy.
The work presents two general capacity laws for finite noisy permutation channels: a structural law for what survives when order is lost, and a resource law for zero-error decoding under local channel constraints. It provides a finite-mmatrix capacity coefficient, demonstrates applicability to diverse channel types, and derives zero-error rate bounds and a convex-optimization frontier for achievable rates under a dependence budget.
GIST is a two-level data selection method for offline RL with verifiable rewards. It selects tokens to supervise using a Gibbs-based mismatch analysis, boosting accuracy across multiple benchmarks while reducing the number of supervised tokens.
SPLIT introduces an exact, Bayes-guided method to allocate test-time compute and verification resources for language-model tasks, outperforming fixed rules across diverse benchmarks; it models verifier behavior, task difficulty, and correlations to optimize sample and budget choices with near-optimal regret.
The documentation explains UMAP 0.6’s compatibility, Adam, and momentum layout optimizers, along with how initialization and compatibility settings affect results. It also describes a warmed-up MNIST timing comparison and how to preserve behavior from earlier UMAP releases.
Translated title. Original: NVIDIAの「長文コンテキスト不要」論文
The article explains an NVIDIA method that compresses agent history and searches long documents selectively to reduce token use. Small tests show major token savings, but repeated searches can slow responses and selective reading can miss the broader meaning of a document.
The paper broadens a fixed-point bundle method to product-of-simplex domains to compute Nash equilibria in general-sum games, showing a geometric equivalence to an equivalent simplex VI which yields global linear convergence; it reports 1e-5 precision on 5,600 random instances and provides code and data online.
The paper argues for a single, stationary intrinsic reward for unsupervised reinforcement learning based on an Expected Free Energy objective. This reward uses information gain to drive exploration in unresolved dynamics and shifts to an aleatoric penalty as dynamics become explained, with a windowed, stationary Bellman framework and nonparametric concentration guarantees.
The paper reports that in multi-agent LLM setups, the model with more parameters tends to dominate the final consensus, absorbing the views of smaller models. It models agents as weighted particles on a simplex, where larger models’ weights yield stronger in-context learning and greater influence, offering a way to interpret complex multi-agent behavior.
A guide to weight-space experiments using PPO on four tasks with anonymized checkpoints, a shared actor-critic network, and a harness that prompts models with raw weights; it details tasks, observations, actions, training/evaluation steps, and anonymization/read-harness workflows.