Source: Teacher-Student Curriculum Learning (arXiv:1707.00183), as a domain Mixer.

Motivation

Sample domains in proportion to learning progress — the slope of each domain's reward curve — rather than its reward level. This is a distinct signal axis from reward_gap (reward level) and dump_ucb (|advantage|): TSCL keys on the rate of change. Both rising (still learning) and falling (forgetting) reward are informative, so magnitude of the slope drives sampling.

Rule

Per domain d, maintain a sliding window of mean rewards and fit a least-squares slope slope_d = dR/dstep. Then:

$$p_d = \mathrm{softmax}\!\left(\frac{|slope_d|}{T}\right), \quad \text{with a floor}$$

• |slope| (default): sample domains that are actively changing (progress or forgetting).
• signed=True: use max(slope, 0) to sample only improving domains.

In DataFlex-RL

trainer:
  v1:
    trainer_mode: dataflex_mix_sync
    sampler:
      custom_sampler: {path: pkg://dataflex_verl.replay_buffer, name: DataFlexMixReplayBuffer}
dataflex:
  mechanism: mix
  domains: [math, code, logic]
  scorer:   {name: reward_difficulty}
  actuator: {name: tscl, params: {temperature: 1.0, floor: 0.05, signed: false}}
  warmup_step: 5
  update_step: 5
  window: 50

The mix trainer computes each domain's reward slope from DomainStatsTracker.slope() (least-squares over the sliding window) and passes it to the mixer.

Signal / rule / granularity

Property	Value
Signal	per-domain reward slope (learning progress)
Rule	`softmax(
Granularity	domain
Needs groups	no
Extra forward pass	no

Needs ≥3 heterogeneous domains

Like all mixers, TSCL adds little with 2 domains (reduces to one ratio knob). It pays off with ≥3 domains of differing, evolving difficulty. Needs a warmup to accumulate enough window points for a stable slope.