For decades, ABR has kept two kinds of intelligence apart. Neural policies learn rich behaviors yet forget them when the environment changes; rules never learn, and never forget. We dissolve this boundary, placing the rule inside the network rather than beside it. But a union must first be testable, and ABR cannot measure what its policies learn or forget. Its yardstick, bandwidth statistics, misleads: identical statistics can hide entirely different outcomes; different statistics, similar ones. We therefore propose Texture-Aware Generalization Evaluation, which judges a policy over its whole training trajectory, on traces whose texture is reported, not assumed.
What breaks a policy is invisible to statistics, yet rules pass through it untouched, reasoning from physics and owing the data nothing. Neuro-Symbolic Manifold Alignment (NSMA) embeds rule decisions as anchors inside the latent space of the neural policy, which keeps learning where learning pays and no longer forgets what rules have always known. Generalization cannot be argued, only survived. Raised on 3G traces alone and released without fine-tuning, NSMA outperforms every state-of-the-art baseline across eight unseen datasets spanning 4G, 5G, and WiFi, and a real-world player. Probing its latent space returns the answer the design promised.
Every conclusion in this field comes from the same kind of evaluation:
testing across trace datasets that differ in name, origin, and statistics.
We questioned this practice, and it failed in both directions.
fcc differs from the training set in source and statistics,
yet agents keep their full advantage on it. train_uncorr
shares every summary statistic with the training set — it is merely a
reshuffled copy — yet on it the advantage disappears. The dataset that
looks different is harmless; the dataset that looks the same is fatal.
We call this the Dataset-Identity Bias, and the hidden variable
that actually decides difficulty, Trace Texture: the intrinsic
character of a trace beyond its summary statistics.
Dataset identity does not predict generalization. Top: statistical
profiles of six trace environments. Bottom: QoE of learning-based
agents versus RobustMPC (gray band) — fcc is statistically far
from train yet performs like it, while train_uncorr
raises no statistical alarm yet collapses.
The advantage of learned policies vanishes under texture shift alone — with every marginal statistic held fixed.
An honest protocol. We codify the repair as Texture-Aware Generalization Evaluation (TAGE): (1) trajectory-based assessment — evaluate the whole sequence of checkpoints saved during training, not a single checkpoint cherry-picked for its performance on the training distribution; and (2) trace characterization — report each environment's marginal statistics and temporal properties. A trajectory cannot be cherry-picked, and with every environment's temporal character on record, "diverse" is no longer a name but a claim that can be checked. Every evaluation in the paper, our own method's included, follows this protocol.
Prior hybrids place the rule outside the network — before it as a teacher (imitation, warm-starting), after it as a guard (switching, safety constraints), or beside it as a colleague (two-stage designs). In every role the rule remains a second party. NSMA removes the assumption: a rule never reads a trace — it reasons from buffer dynamics, throughput conservatism, and lookahead constraints — so a texture shift can take nothing from it. NSMA lifts the rule's decision into a trainable logic embedding G and places it inside the representation itself, jointly trained and differentiable end to end.
A direct addition of the logic embedding G to the feature embedding. Even this minimal form already does what no prior hybrid did: the rule's signal enters the representation end-to-end, and it already beats every baseline (QoE 0.831).
Hyper-Connection Routing replaces the fixed addition with a routed, state-dependent one: the representation widens into parallel streams, the state decides the routing, G is fused before the transformation, and the streams settle into a consolidated state. The network learns; the anchor holds (QoE 0.848).
Under the TAGE protocol, RobustMPC sets the level that learning-based methods have struggled to pass: Pensieve, Merina, and PA-MoE all converge just below it (0.782–0.801 vs 0.803). NSMA-HCR leads throughout training and clears it, converging above 0.85 — with no cold start (it opens above 0.80, because the anchor carries what the rules know into the very first decision), and at no cost in safety: it reaches the bitrate of aggressive RL policies while keeping rebuffering and smoothness at rule-based levels.
Which half carries the weight? Two controlled sweeps answer it. Fixing the architecture and varying the rule barely moves the result (0.80–0.82); fixing the rule and varying the architecture swings it from strong gains to outright negative transfer. The value of a logic prior is real but conditional — the condition is an architecture able to absorb it, and HCR is our answer to that condition.
Does HCR truly anchor the representation, or merely add a useful input? We corrupt the logic embedding G with white noise at three strengths and blend the corrupted and the real embedding, sweeping from pure noise to pure logic. Mild damage barely hurts; at every noise strength QoE rises steadily and smoothly as the real embedding is restored — exactly the behavior an anchor predicts.
The decisive test runs on states collected from real playback. With the true G, a UMAP projection of the actor's internal features is cleanly ordered along the bitrate ladder. Shuffling G across states keeps the injected vectors — and hence the information content — identical, destroying only the state↔rule correspondence: the ordering collapses. A 10-nearest-neighbor probe puts numbers on it: with the true G, rule-action decodability rises from the raw-state level of 0.433 to 0.872. The structure comes from the pairing, not from adding a vector.
Foreign networks. Trained on 3G traces alone, the policy is
deployed with no fine-tuning on Lumos4G (mean 40.4 Mbps), Lumos5G
(522.9 Mbps), and SolisWi-Fi (24.2 Mbps) — bandwidth
scales up to two orders of magnitude beyond training. Existing learned
policies flatten out far below the rule-based methods; NSMA-HCR is the
only learned policy that climbs close to the rule-based line on every one
of them, and the one method that ends clearly above RobustMPC on
train_uncorr.
Real hardware. Each method runs 27 sessions of real HLS playback
over 3 seeds, on an unmodified hls.js player in headless
Chromium, with bandwidth throttled in real time by the traces. Real
playback proves harsher than any simulation: every baseline — even
RobustMPC — ends with negative QoE. Only the two NSMA variants stay
positive, and NSMA-HCR shows the smallest sim-to-real gap: it wins on real
hardware, and its advantage is not tied to the simulator.
@inproceedings{he2026nsma,
author = {He, Zhiqiang and Liu, Zhi},
title = {{NSMA}: Neuro-Symbolic Manifold Alignment for Generalizable
Adaptive Bitrate Streaming under Texture Shift},
year = {2026},
}