The paper investigates how identical model weights can yield different next‑token distributions when run by different serving programs, due to variations in precision, kernel implementation, and batching. It introduces the concept of kernel contracts that bound how much a serving kernel may deviate from a reference kernel on each traffic slice, and proves non‑asymptotic bounds linking logit differences to divergence metrics, reward changes, policy‑gradient bias, and greedy‑token flips. Experiments on Qwen2.5‑7B‑Instruct confirm that these bounds hold across prompts and that precision or engine changes can alter token outputs, especially on undecided prompts, highlighting the need for per‑slice contract checks.
arXiv:2606. 07581v1 Announce Type: cross Abstract: A modern post-training pipeline often writes one symbol for its policy, pi_theta, while evaluating it through two different programs: a training kernel optimized for autograd and an inference kernel optimized for low-precision, fused, dynamically batched serving.
arXiv:2608.29296v1 Announce Type: cross Abstract: Larger batches reduce the variance of stochastic gradients per update and are therefore often expected to accelerate training. Yet whether this stati...
ScaleRLT introduces a hybrid learner log‑probability method that reconciles differences between dense and deferred rollout computations in looped language models like Ouro. By using a stop‑gradient replay of the rollout schedule, the approach achieves bit‑for‑bit reproducibility across FP32 and bf16 real‑weight cells while maintaining the same gradient accuracy as dense implementations. The technique boosts rollout throughput by up to 1.56× on a single GPU and significantly speeds up multi‑node GRPO updates, leading to substantial accuracy gains on GSM8K in short training runs.
The paper presents an end‑to‑end MXFP8 reinforcement learning implementation on Blackwell GPUs, evaluating its benefits and limitations with Qwen3-8B and Qwen3-30B-A3B models. It finds that matching FP8 precision between sampler and learner improves consistency, that speed gains appear only for large micro‑batches, and that several weight‑synchronization correctness conditions can fail silently. The study also introduces seven explicit checks to detect such failures during training.
Asynchronous reinforcement learning improves throughput by decoupling rollout generation from optimization, but staleness is an inevitable byproduct compounded by policy lag, engine delays, and mixture-of-experts routing. From a trust-region perspective, this mismatch is critical: training-inference divergence governs approximation error in finite-horizon bounds, whereas PPO clipping only gates sampled outward updates, acting as a sampled surrogate rather than a full-policy constraint.
arXiv:2607. 18722v1 Announce Type: new Abstract: Asynchronous reinforcement learning improves throughput by decoupling rollout generation from optimization, but staleness is an inevitable byproduct compounded by policy lag, engine delays, and mixture-of-experts routing.
HeadWiseKV is a training‑free framework that compresses the residual global key–value caches of hybrid long‑context language models by assigning each physical KV head a static, multilevel history window. It formulates cache allocation as a restricted operational rate–distortion problem and uses the SeqCalib algorithm to generate per‑head residency policies that account for interactions across layers. In evaluations on four hybrid models, HeadWiseKV preserves near‑full‑KV quality while reducing peak device memory usage by 8.59% at a 112K context length and extending the largest verified context from 114K to 161K.
arXiv:2610.09519v1 Announce Type: new Abstract: Production machine-learning models are derived artifacts of time-bounded training snapshots: a deployed model is a materialized view over a training cu...
arXiv:2607. 14427v1 Announce Type: new Abstract: A depth-recurrent transformer applies a weight-tied core a variable number of times, and prior work has shown that training with a randomized recursion count yields one checkpoint usable across a range of inference depths.
arXiv:2606. 05597v3 Announce Type: replace Abstract: Training vision-language web agents with multi-step RL is compute-intensive, with two dominant forms of inefficiency: idle GPUs in synchronous RL, and trajectories that use more steps and tokens than necessary.
arXiv:2607. 14952v1 Announce Type: new Abstract: A growing gap separates inference context lengths from RL post-training: inference systems are approaching million-token contexts, while post-training workloads often remain at 256K tokens or below and rely on length generalization at deployment.
arXiv:2607. 22186v2 Announce Type: replace Abstract: Asynchronous reinforcement learning (RL) accelerates large language model (LLM) post-training by overlapping rollout generation with policy optimization, but the resulting stale, off-policy data can destabilize optimization and ultimately cause policy collapse.