Every data-scheduling method in DataFlex-RL reduces to a signal (a scalar/tensor computed per token / sample / group / domain) fed to a rule (select / reweight / mix). This page is the single place to see, across all three mechanisms, what signal each algorithm keys on and where it comes from.

The research frontier is mostly about new signals — the rules (top-k, softmax, band, UCB, threshold, max-variance) are shared and interchangeable.

Master table

Algorithm	Mechanism	Signal	Granularity	Source (arXiv)
threshold_band	select	group solve-rate / any score in a band	prompt/group	DAPO 2503.14476
difficulty_filtering	select	group pass-rate, keep 0.2–0.8 band	prompt/group	2504.03380
gfpo	select	response length + reward (conciseness / token-efficiency), top-k per group	response(group)	2508.09726
pods_maxvar (max_variance)	select	per-response reward, max-variance subset	response(group)	2504.13818
topk_fraction	select	per-sample |advantage|, keep top fraction	prompt	generic (PER-style)
advantage_reweight	reweight	per-token π_θ (probability), damp low-prob tokens	token	2505.12929
per_advantage	reweight	per-sample |advantage| (w=|A|^α)	prompt	PER 1511.05952
softmax	reweight	any per-sample score (softmax(s/T))	prompt	generic baseline
difficulty_band	reweight	per-seq reward quantile band	prompt	generic (cf. ODSW)
dump_ucb	mix	per-domain |advantage| + UCB	domain	2504.09710
tscl	mix	per-domain reward slope (learning progress)	domain	1707.00183
reward_gap	mix	per-domain mean reward (favor lagging)	domain	generic (cf. DoReMi)
static	mix	fixed proportions	domain	baseline

Signals grouped by axis

• Reward level / pass-rate — difficulty_filtering, pods_maxvar, difficulty_band, reward_gap, threshold_band.
• |advantage| — topk_fraction, per_advantage, softmax (as used), dump_ucb.
• Reward slope / learning progress (rate of change, a distinct axis from reward level) — tscl.
• Token probability — advantage_reweight.
• Length / token-efficiency — gfpo.

On provenance (honest note)

Most algorithms map a published method's core idea onto DataFlex's Scorer→Actuator abstraction; a few (topk_fraction, softmax, difficulty_band, reward_gap, static) are generic mechanism knobs the abstraction provides, not faithful reproductions of a single paper. Where a paper is named, the implementation captures its signal + rule, not necessarily every training trick in the original. See the survey (Awesome-RL-OPD-Data) for evidence grades and full sourcing.