Research Briefing — June 16, 2026

Hyperball is an optimizer wrapper that fixes the Frobenius norms of weight matrices and their updates to constants, restoring the scaling gains of matrix optimizers like Muon over AdamW. It achieves 20 to 30 percent token-equivalent speedups and improves learning-rate transfer on Qwen3-style models up to 1.2B parameters.

A large empirical study of over 100 biological reasoning models examines how continued pre-training, SFT, and RL each shape in-domain versus out-of-domain generalization across genomics, transcriptomics, and proteins. It identifies when post-training improves performance versus inducing over-specialization.

Formalizes diversity collapse in RLVR (Pass@1 improves while Pass@k degrades) through the lens of overtraining, showing that once a problem's reward contribution saturates further updates just concentrate probability mass. Provides theoretical grounding for a known RLVR pathology.

Constitutional Value Potentials learn scalar potentials from hidden states representing internal pressure to preserve each value, with their signed difference forming a priority margin readable from activations. Aims to detect and steer how models arbitrate value conflicts internally.

Argues that visual representation learning should remove remaining inductive biases like augmentations and masking, proposing a temporal-difference approach that learns from video without strong assumptions. Experiments confirm that the optimal strength of inductive biases diminishes with scale.

This paper introduces a fast online estimator of preconditioned Hessian curvature using warm-started power iteration, then uses it for an architecture warm-up that stabilizes billion-parameter Transformer training. It connects Edge of Stability theory to practical large-scale training.

This paper finds that language models internally track the value of their current reasoning trajectory along a learnable value axis, distinguishing high vs low confidence, backtracking, and correct vs corrupted code. Steering along this axis causally controls self-correction and exploration behavior.

This work extends Pearl's structural causal models to relational settings where objects and relations vary, deriving identifiability conditions and relational causal graphs for generalizing to unseen object combinations. Co-authored by Elias Bareinboim, it offers rigorous theoretical foundations for combinatorial causal reasoning.