Research Briefing — August 12, 2026

A theoretical lower bound of Omega(T^-1.9319) for last-iterate convergence of gradient descent with predetermined nonnegative stepsize schedules in smooth convex optimization. The result shows that stepsize schedules alone cannot accelerate plain GD beyond certain rates.

Empirical evaluation of machine-unlearning robustness on WMDP-Bio checkpoints across six methods (RMU, ILU-RMU, NPO, GradDiff, NPO-ILU, IDK-AP). Tests refusal-direction ablation, forget-set representation-targeted ablation (extending Arditi & Chughtai), and unrelated SFT, finding broad recoverability of supposedly unlearned knowledge with the right probe.

Ouroboros is a self-developing coding agent whose harness, prompts, context assembly, and core implementation improve through reviewed commits that recursively become the runtime for future work. It reports strong results on Terminal-Bench 2.1 (86.74%) and OSWorld-Verified (90.69%) using Opus 5, plus the best reported CL-Bench reward after a five-rollout campaign. The novelty is sustained, reviewed self-modification rather than one-shot self-edit.

Combines a goal-conditioned Joint-Embedding Predictive Architecture with World-Action Models to explicitly predict task-relevant semantic stages in robot manipulation, achieving 91.42% success on clean and 89.08% on randomized RoboTwin 2.0 tasks.

Constructs an interactive PCP protocol enabling polynomial-time verification of approximate consistency for a probabilistic predictor specified by probability circuits paired with a confidence circuit. The work is motivated by AI safety settings where verifying honesty about risk predictions matters.

SWE-Bench ProMax is a curated multilingual benchmark of large-scale, real-world code refactoring tasks designed to stress-test AI coding agents beyond the original English-centric SWE-Bench. It exposes substantial headroom that current agents have not closed, making it a useful reality check amid rapid coding-agent progress. The curation rigor is the main contribution.

Model Discovery Agent couples an LLM as a proposer of causal structures with SMC for parameter/structure posteriors, simulation-based inference, and value-of-information experiment design to discover mechanistic world models from few interventions.

REDAgentBench is an executable red-teaming framework that separates exposure, execution, observation, and adjudication in LLM agent safety evaluation, running attacks in isolated service sandboxes and verifying harmful effects from service receipts.

XCoT-VLA replaces verbose natural-language Chain-of-Thought in autonomous driving VLA models with compact executable CoT tokens learned from automatically constructed Reason-Action supervision, conditioning trajectory queries via shared multimodal self-attention.

A case study documenting how AI research agents were used to tighten known bounds on the Grothendieck constant K_G, yielding new ranges. The paper analyzes strengths and weaknesses of AI assistance in mathematical research, with authorial expertise from Pravesh Kothari and Raghu Meka.

OasisKV extends LLM inference KV cache beyond GPU HBM by storing full caches in lower memory tiers and prefetching only context-relevant entries into HBM using a speculative-decoding-style lookahead predictor. Aims to break the memory wall for very long-context inference without accuracy loss. Relevant for serving long-context models cheaply.