FourSight targets concentrated errors to stabilize four-bit AI training

Why cheaper arithmetic can disrupt learning

Reinforcement-learning post-training alternates between generating responses and updating a model using rewards. These stages often run in separate software engines, each optimized for its own workload. Even with identical model weights, numerical differences can make the engines assign different probabilities to the same token.

The FourSight preprint, by Yining Wei and colleagues, investigates how NVFP4—a format using four-bit weights and activations with additional scaling—amplifies this training–inference mismatch. In their experiments, native FP4 training can improve initially, then collapse and recover only partly, despite an existing token-level correction.

The authors identify two related problems: generated responses deteriorate under low precision, and mismatched responses can produce destabilizing policy updates. Their proposal addresses both without returning the whole computation to higher precision.

Finding the concentrated failures

In a diagnostic experiment with Qwen3-1.7B on GSM8K, severe mismatch clustered within particular responses. Tokens that the generation engine produced could receive extremely low probability from the training engine. The authors associate these cases with repetitive symbols, garbled fragments and abrupt language shifts.

This concentration matters because the group-relative training algorithms studied assign a shared advantage—a reward-based learning signal—to every token in a response. Suppressing only individual mismatched tokens can therefore leave other parts of a derailed response contributing updates.

A separate Qwen3-8B experiment examined whether degraded performance reflected damaged model weights. Switching a degraded FP4 checkpoint’s generation precision to BF16, without changing its weights, recovered approximately 20 percentage points of GSM8K test accuracy. The authors interpret this as evidence that increased sensitivity to quantization, rather than irreversible weight corruption, drove much of that degradation.

They also found large quantization errors concentrated in a small number of activation channels. NVFP4 shares a scale across blocks of 16 values, so one large outlier can make the numerical representation coarser for its neighbors.

Two selective interventions

FourSight first filters entire generated sequences. It marks tokens whose training probability falls below a small threshold, then excludes a sequence from policy updates if it contains too many such tokens overall or a sufficiently long consecutive run. Isolated low-probability tokens need not disqualify an otherwise useful response.

Filtering sets the sequence’s advantage to zero. Importantly, the method retains the original group-size denominator rather than renormalizing around surviving sequences, avoiding an unintended amplification of their gradient contributions.

The second intervention protects selected channels using BF16, a higher-precision format, while retaining native NVFP4 elsewhere. The experiments generally protect the 64 highest-variance input channels per projection, reselecting them each training step using previously accumulated statistics.

A fused kernel called SpliceGEMM combines the four-bit computation and BF16 correction. It uses different execution paths for larger token batches and small-batch decoding, reducing the overhead of running a separate correction operation.

What the experiments show

For Qwen3-8B trained with DAPO on DAPO-Math-17k, FourSight first reached 20% AIME24 validation accuracy after 150 steps with Megatron-LM training and 250 with FSDP. Relative to QaRL, those results represent 2.60-fold and 1.76-fold improvements in steps to that target—not universal wall-clock speedups.

Deployment evaluation used FP4 inference for every final checkpoint, including models trained in BF16. Methods with precision compensation retained it. Under those conditions, FourSight matched or exceeded the quantized comparison methods on nine of ten Qwen3-8B benchmark–engine combinations. For Qwen3-1.7B, it exceeded the BF16-trained baseline by 12.7–16.4 percentage points across four benchmarks. That comparison does not mean it outperformed BF16 inference: the BF16-trained baseline was also converted to FP4 for evaluation.

Separate timing experiments, using matched response-length distributions, covered models up to Nemotron3-Super-120B-A12B. FourSight was at most 2.3% slower per step than uncompensated FP4 in those measurements, and slightly faster in two larger-model configurations.

Where the evidence stops

This is a preprint, not established independent validation. The detailed training-quality experiments concern two Qwen3 sizes and mathematical-reasoning workloads; the 120B-scale result concerns system performance, not equivalent evidence of learning quality at that scale.

The channel-selection theory is conditional, too. It minimizes a relaxed error bound using second moments; variance provides the same selection under specified mean conditions. The analysis also assumes nonsaturating normal block scales and an exact BF16 correction branch.

Ablations support combining filtering and compensation, but their relative benefits differ between training engines, and the combination does not win every individual benchmark. The evidence supports a targeted engineering approach within the tested settings, rather than a guarantee for every model or task.