When FP8 Training Stops Punishing Bad Answers

Featured

01 Training: FP8 RL Can Forget Bad Answers

Full-pipeline FP8 RL may fail for a subtler reason than low precision: some bad answers quietly escape punishment. The paper finds that accumulated quantization noise distorts importance ratios, which measure how far the new policy has moved from the old one.

Negative-advantage tokens then cross the clipping boundary more often, incorrectly reducing their gradients to zero. When the model produces gibberish or other failures, the corrective signal disappears. Those failures accumulate until training entropy suddenly spikes.

Calibrated Clipping aligns the FP8 lower clipping boundary with a lower quantile from the BF16 distribution. It adjusts the upper boundary accordingly. The abstract reports no entropy spikes across GRPO, DAPO, models from 8B to 32B, and several FP8 scaling granularities. Performance returned close to the BF16 baseline.

Training teams should monitor more than final scores. Check whether negative-advantage tokens continue receiving gradients, or acceleration may silently disable the system’s ability to punish bad behavior.

Key takeaways:

• Full-pipeline FP8 RL may fail through lost corrective signals, not just lower average precision.
• Monitor entropy, abnormal outputs, and gradients for negative-advantage tokens during quantized training.
• Calibrated Clipping may stabilize full-pipeline FP8 RL, though its exact overhead requires the full paper.

Source: Towards Full Pipeline FP8 Reinforcement Learning for LLMs

02 Training: Balance Supervision Across Teaching Capabilities

OmniEdu organizes educational supervision around four capabilities: subject knowledge, curriculum alignment, diagnostic reasoning, and instructional action with scaffolding. It consolidates more than 100 sources into nearly 70,000 samples through semantic review, quality scoring, and diversity filtering.

This balanced training improved curriculum understanding, K-12 problem-solving, and tutoring benchmarks across 4B, 9B, and 27B models. The results suggest that one data pipeline can teach both problem-solving and guided instruction.

Educational model teams may gain more from covering the full teaching workflow than simply adding data. The evidence still comes mainly from standardized benchmarks. It does not measure long-term learning, harmful interventions, or differences between students in real classrooms.

Key takeaways:

• Organize educational data around the capability chain, not only by source or question type.
• Balanced training improved both problem-solving and tutoring, but the abstract does not isolate each supervision category’s contribution.
• Benchmark gains support the training direction but do not establish real teaching effectiveness.

Source: OmniEdu: Open Foundation Models for Learning and Teaching

03 Image Gen: Remove Background Tokens First

A unified sequence for 2K multi-view texture generation can exceed 212,000 tokens. Most wasted compute comes from background tokens and sparse interactions among foreground tokens.

UltraTex treats these as separate problems. It drops background tokens entirely and applies block-sparse attention within the foreground. The method also modifies VAE decoding so the decoder explicitly recognizes foreground regions, reducing reconstruction artifacts.

On common samples from its dataset, training ran 20.6–91.1 times faster than the selected baseline. End-to-end inference improved by 22.3–74.6 times. These figures should not be generalized across models, assets, or production environments without testing.

For high-resolution generation systems, first identify each source of redundancy. Then modify attention and decoding separately.

Key takeaways:

• Separate information that can be deleted from interactions that must remain but can use sparse computation.
• Removing background tokens is insufficient; the decoding path must also adapt to control artifacts.
• The reported speedups depend on a specific baseline and in-dataset samples. Retest with your own assets before production use.

Source: UltraTex: Unleashing 2K Multi-View Diffusion for 3D Texturing