Alpha-v2.6 extends the v2.4 distributional line with a multi-horizon QR-TD3 critic. The critic is reorganised around two orthogonal axes that do not interact — they simply multiply: num_quantiles (the granularity of the return distribution) and nstep_horizons (how many distinct n-step bootstrap targets the same Q* is estimated against).
Result. One DiverseCritic per symbol now emits a (B, N, 4, 32) tensor — four return distributions (one per horizon), each described by 32 quantiles. 4 × 32 = 128 values per symbol per critic, versus a single 32-quantile distribution before.