A preprint finds that conditional flow matching, a method that trains a model to predict motion along a probability path, can make endpoint likelihood estimates much more accurate without making them exact. The same analysis shows why the distinction matters: during on-policy training, a flow-matching ratio can produce high reported reward while remaining a poor match for the clean likelihood ratio.
The study’s central result is a mathematical condition, not a blanket endorsement of the shortcut. For the linear-Gaussian path examined in the theory, endpoint negative log-likelihood, or NLL, can be written as an entropy term, a weighted pointwise CFM objective, and two residual terms that capture what the shortcut leaves out. The CFM-only estimate is exact only when those residuals cancel.
A better estimate, with a boundary problem
The study then tests a score-calibrated weighting scheme in a fixed-target, or off-policy, setting. At the population optimum for that fixed target, the weighting removes the interior residual. A positive-time boundary term can still remain, along with numerical error, so calibration improves the identity without guaranteeing exactness.
In the reported summaries, score calibration cut the mean absolute error of CFM-only NLL estimates from 0.800 to 0.223 in one dimension, from 1.294 to 0.377 in two dimensions, and from 19.402 to 3.012 in 32 dimensions. The full decomposition, which includes the residual terms, was more accurate than the CFM-only estimate across those off-policy comparisons.
Those numbers describe fixed-target summaries, not a promise about a model while it is still learning. During optimization, neither ordinary nor score-calibrated CFM has a general pointwise NLL identity because the needed pointwise calibration is not enforced. A weighting rule that works at the population optimum does not automatically make every training-time estimate faithful.
Why reward can look good while ratios go wrong
The gap becomes sharper in the on-policy setting, where ratios are evaluated on samples relevant to the update. There, the CFM-only old/new ratio matches the clean endpoint ratio only if the relative interior and boundary residuals vanish on those samples. A stated Gaussian counterexample makes the point directly: the CFM log-ratio can be nonzero even when the clean endpoint log-ratio is zero.
The analysis also shows that regularized optimization can leave this mismatch in place. After optimizing a CFM-ratio surrogate, the clean endpoint log-ratio can remain zero while the CFM-only ratio remains nonzero. Success on the surrogate’s objective is therefore not, by itself, evidence that the underlying likelihood ratio has become accurate.
Representative runs still reached rewards of 0.995, 0.981, and 0.910 in one, two, and 32 dimensions. But the corresponding log-ratio mean absolute errors were 7.900, 1.570, and 1382.955. The result is a warning for anyone reading reward as a broad measure of ratio quality: a useful optimization direction can coexist with substantial bias.
The shortcut becomes fragile as dimension rises
The high-dimensional comparison produced a similar split. Ordinary CFM-ratio training passed the pre-specified 0.90 reward gate in every tested dimension from 4D through 32D, with final rewards of 0.997, 0.995, 0.991, and 0.910. The score-calibrated version passed only in 4D, at 0.998, then fell to 0.643, 0.481, and 0.209 in 8D, 16D, and 32D.
The reported high-dimensional comparison supports a local optimization claim, not a general guarantee. The exact result is derived for the linear-Gaussian path used in the theory, while the experiments cover synthetic distributions with known endpoint densities and a raw-image stability test. Those experiments do not establish that CFM-only objectives are generally exact outside the stated setting.
Stability controls are not accuracy guarantees
The study’s control sweeps associated the freshest exponential moving average, or EMA, reference with the most consistent passage of the reward gate. Ratio clipping was associated with the most passing configurations when reference updates were slower. These were empirical associations, often from single-seed configuration studies, and did not establish causality or local ratio accuracy.
Raw-image tests were used only to ask whether such optimization mechanisms transfer to image-valued models. The study did not use them to assess likelihood or decomposition claims because raw-pixel CNF likelihood calculations and pointwise audits were inaccurate and prohibitively expensive at image dimensions. The tests therefore provide a stability check, not an exact likelihood audit.
In the raw-image sweep, one setting had 2 of 2 successful configurations with EMA set to 0 and 0 of 2 at EMA values of 0.9 and 1. Another had 6 of 8 successful configurations at EMA 0 and 1 of 8 at EMA 1, using a 0.90 reward-success gate. This pattern is consistent with the study’s stability associations, but it does not establish causality or local ratio accuracy.
The preprint therefore draws a careful line between exactness and usefulness. Its decomposition says a CFM shortcut becomes an exact endpoint likelihood calculation only through residual cancellation. Its experiments suggest that score calibration can improve fixed-target estimates, while an on-policy CFM-ratio surrogate can still help reward optimization despite biased ratios. Broader probability paths, better control of on-policy residuals, and possible applications to likelihood-based post-training for flow-matching models remain open questions.
The supplied document is labeled a preprint, and no funding information is reported in the supplied text or metadata.
Paper data and sources
Original title: When Can Conditional Flow Matching Replace Pointwise Negative Log-Likelihood?
Authors: Yansen Han, Hongxin Sun, Tao Lin
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text