Recorded on 2026-08-26 from the user-reported remote training and infer_turn_head.py output. This is a single-run diagnostic result, not yet a multi-seed paper number.

Experiment contract

Dataset: EasyCom.
Test split: confirmed as EgoSpeak official Sessions 3, 2, and 1. Training is recorded as the script's default official Sessions 4--12; verify this against the remote checkpoint JSON because the supplied excerpt starts at epoch 24.
Backbone: frozen Qwen2.5-Omni-3B Thinker.
Streaming mode: open-stream replay with an 8-second causal-memory budget.
Input sampling: 2 FPS video and raw audio.
Prediction rate: 5 Hz, with ten future horizons from 0.2 to 2.0 seconds.
Readout: the query states from the final four Thinker layers, ordered oldest to newest, followed by a learned global softmax layer mixture and the same 10 x 3 anticipation MLP.
Training: head and four scalar layer logits only; Thinker remains frozen.
Evaluator: EgoSpeak full-timeline per-frame AP/mAP.

The remote artifact is named easycom_frozen_3b_8s_last4_learned_mix_50ep.pth, but the supplied log and the user report both show that the actual run ended at epoch 25/25. The 50ep suffix is therefore a naming error and must not be used as provenance.

Summary

All values below are percentages. The comparison baseline is the previously recorded frozen 3B, 8-second, open-stream, single-final-layer run evaluated on the same official test sessions.

Readout	0.2s full mAP	Full mean mAP	Delta at 0.2s	Delta mean
Single final-layer query	71.99768	58.51937	--	--
Last-4 learned mixture	73.45085	60.53423	+1.45317	+2.01486

The last-4 readout improves every evaluated horizon. The largest gain is +2.41084 at 1.0 seconds; the smallest is still a positive +1.45317 at 0.2 seconds.

Per-horizon full-timeline results

Horizon	Last-4 mAP	Single-layer mAP	Delta	Background AP	Wearer AP	Others AP
0.2s	73.45085	71.99768	+1.45317	56.69385	78.35111	85.30760
0.4s	68.58645	66.74289	+1.84356	49.77891	73.48058	82.49988
0.6s	64.44213	62.38972	+2.05241	44.51563	68.39317	80.41758
0.8s	61.34852	59.12229	+2.22623	40.98328	64.20310	78.85917
1.0s	59.12464	56.71380	+2.41084	38.99833	60.75541	77.62018
1.2s	57.53667	55.36300	+2.17367	37.76815	58.15110	76.69076
1.4s	56.43710	54.21297	+2.22413	37.22701	56.12642	75.95786
1.6s	55.42035	53.43221	+1.98814	36.72606	54.14045	75.39454
1.8s	54.76334	52.84703	+1.91631	36.45931	52.80226	75.02845
2.0s	54.23227	52.37207	+1.86020	36.45292	51.56488	74.67902
Mean	60.53423	58.51937	+2.01486	41.56035	61.79685	78.24550

Averaged over horizons, the AP gains relative to the single-layer baseline are +1.92941 for background, +2.88478 for camera-wearer speaking, and +1.23041 for other-people speaking. The wearer class contributes the largest improvement, which is encouraging for the short-horizon turn-taking objective.

Learned layer mixture

Layer order	Weight
Oldest of selected four	0.246384
Second	0.248327
Third	0.241774
Final layer	0.263515

The weights remain close to uniform: normalized entropy is 0.999619, the L1 distance from [0.25, 0.25, 0.25, 0.25] is only 0.02703, and the largest weight exceeds the smallest by only 0.021741. Consequently, this experiment supports multi-layer query readout, but does not yet show that learning the mixture is better than a fixed last-4 mean.

Interpretation and next controls

This is the strongest recorded frozen-Omni EasyCom result in the project so far: 73.45085 at 0.2 seconds and 60.53423 mean mAP.
The gain occurs at all horizons and all three classes, rather than coming from a single AP outlier.
Because the learned weights are nearly uniform, the first required ablation is a fixed uniform last-4 mean with the same feature files, head recipe, epoch count, and seed.
Then compare each of the four individual query layers. This separates layer ensembling from access to a specifically better intermediate layer.
Repeat the single-layer, fixed-mean, and learned-mixture variants with at least three identical seeds before using the gain as a paper result.
The supplied output contains no strict causal-valid table. Archive the remote metrics.json or rerun the evaluator before the publication table.
The single-layer comparison came from an earlier run; its random seed and head-training duration must be matched in the fixed-mean ablation.

Provenance

Remote feature root: /data1/yfzx/wangx/Datasets/easycom/omni_features/open_stream_2fps_8s_last4_stack
Remote checkpoint: checkpoints/easycom_frozen_3b_8s_last4_learned_mix_50ep.pth
Remote result directory: results/easycom_frozen_3b_8s_last4_learned_mix_50ep
Final training loss: 0.822849.
Raw console excerpt: raw_console.txt.
Machine-readable values: summary.json and metrics.csv.
Raw remote metrics.json archived locally: no.
Exact training-session list verified from the remote checkpoint JSON: no.
Multi-seed: no.