Yun-Yen
Chuang

One schedule for every sentence

Standard S2S-diffusion (DiffuSeq) adds noise on a fixed √-schedule. A trivial paraphrase and a hard open-domain reply are corrupted exactly the same way.

But sentences differ in difficulty — non-contextualized noise leaves performance on the table.

A network that reads the sentence

The Scheduler Bψ — a small Seq2Seq model — reads the conditioning sentence wˣ and emits a sequence of Meta-Instructions ι, each labelled True or False.

A ‘skipping’ rule turns those labels into noise: True steps the noise up, False holds it.

Less noise for hard sentences

The result is a per-sentence schedule βˣ that bends away from the fixed baseline. Harder sentences get less noise to preserve signal; easier ones get more to boost diversity.

This is the move that non-contextualized schedulers can't make.

Generate with the scheduled noise

The Exploiter Dθ — the S2S-diffusion model — diffuses and denoises using βˣ, recovering z₀ step by step, then rounds it back into a discrete target sentence ŷ.

Generation quality teaches the scheduler

How much did the exploiter improve? Compare BLEU before and after an update: the meta-reward R_β = r′ − r flows back through a policy gradient to train the scheduler.

The scheduler learns how to noise — never touching the generator's loss directly.

State of the art, plug-and-play

Meta-DiffuB sets a new bar across four Seq2Seq benchmarks, beating prior diffusion models and fine-tuned PLMs.

Better still, the trained scheduler drops into existing models like DiffuSeq as a plug-and-play module — no fine-tuning required.

One network, two jobs

Language GANs train a generator with reinforcement learning. But that single generator must both sample (explore) and learn (exploit) — with the same policy.

Rewards are sparse, so exploration is poor: quality and diversity can't improve together, and the model drifts toward mode collapse.

A dedicated explorer

MetaEx-GAN adds a meta-trained Explorer (the teacher), whose only job is sampling. The Generator (the student) is freed to just learn.

Sample a richer space

The explorer rolls out a diverse batch of candidate sequences — searching parts of the space the generator's own policy would never reach.

Generator learns, discriminator scores

The generator learns from that batch; the Discriminator scores real vs. generated and returns a reward that updates the generator by policy gradient.

The student's progress trains the teacher

The generator's learning effectiveness — how much it improved on the explorer's batch — becomes the meta-reward that updates the explorer's policy.

The teacher learns to explore exactly where the student learns most.

Quality and diversity, together

MetaEx-GAN reaches state-of-the-art NLG, improving sampling quality and diversity at once — without generating more sequences.

It also generalizes to large pre-trained generators like GPT-2.

Not every camera is equally reliable

Multiview detection projects feature maps from many cameras onto one ground plane to beat occlusion. The strong baseline, MVDetr, weights every camera equally.

But occlusion depends on object position and camera angle — equal weighting leaves accuracy on the table.

A 2D network to anchor the truth

A single-view detection-by-tracking network (FairMOT on DLA-34) produces reliable 2D foot points, using tracklets to fill in missed detections under occlusion.

Lift each view to the ground plane

Perspective transformation maps each camera's feature map onto the bird's-eye ground plane; a deformable transformer encodes per-camera BEV features.

Reliability = agreement between 2D and 3D

Project the predicted 3D foot point back to 2D and measure its discrepancy against the 2D detection. Small discrepancy means a trustworthy camera.

Averaged per camera, this becomes a camera-aware attention weight Ac.

Guide learning — only when it matters

Cameras are aggregated by attention-weighted averaging, F = (1/C)·Σ Ac·fc. Because the consistency computation is expensive, a query-based-learning scheduler only steps in when the camera weight order shifts — gated by an entropy change above 0.1.

The 2D network acts as an oracle that schedules when to teach the 3D detector — a scheduler, just like in my generation work.

A new benchmark on both datasets

QMVDet sets state-of-the-art MODA on Wildtrack (93.1%) and MultiviewX (95.1%), beating MVDet, MVDetr and 3DROM.

It even stays robust when some cameras drop out.

Rolling out whole sentences is expensive

To reward a partial sequence, SeqGAN runs Monte-Carlo roll-outs to the end of the sentence at every generation step — computationally heavy.

And using the discriminator's score directly invites mode collapse: rap lyrics that all sound the same.

Roll out meaningful phrases

Phrase Roll-Out segments each line into meaningful phrases with TextRank, ρ′(y) = (p₁…p_T′), instead of arbitrary or full-sentence cuts.

Meaningless fragments are filtered out, which keeps the generated lyric fluent.

Reward to the phrase boundary

Each phrase is rewarded only out to its own end step t′_end — not the end of the whole sentence. The roll-out is far cheaper while staying meaningful.

An attention-LSTM discriminator

The discriminator learns attention alignment weights α = softmax(wᵀ·tanh(H)), capturing both local phrase features and global sentence meaning across variable-length phrases.

Match the weights, not the score

Instead of the raw score D(x,ŷ), the generator is rewarded by the cosine similarity S(α*, α̂) between real and generated attention weights.

This feature-matching signal prevents mode collapse and lifts diversity and originality.

Better lyrics — and a dataset

RapGAN beats SeqGAN, MaliGAN and Ghost Writer on diversity, originality and fluency, validated against human evaluation.

It also releases an open 160,000-song Chinese rap dataset for future research.

Feeds let negativity pile up

Social feeds display posts first-in-first-out. When many support-seeking, negative posts land together, readers absorb them all at once.

Psychologists call it the cost of caring — repeated exposure drives stress and depression.

Label boards, learn embeddings

Posts crawled from a BBS are labelled by source: the Prozac and Hate boards as social-overload, others as normal — balanced to avoid skew.

Words are embedded with Word2vec Skip-Gram.

A document-level detector

CKDGNN stacks word embeddings → CNN filters → K-max pooling at the document level → a GRNN → softmax, scoring each post's overload probability.

It reaches 95.15% accuracy, beating DCNN, CNN-GRNN and LSTM-GRNN.

Flag with a threshold

Every post gets a probability. Anything above the 0.5 threshold is flagged as social-overload — the rest are normal load.

Rearrange to protect the reader

The Social-Overload prevention System re-sorts the feed so no more than three overload posts appear in a row — inserting a calmer post to break the streak.

Reader load stays under the tolerance line.

A measurably calmer feed

Detection hits 95.15% accuracy, and after rearrangement 75% of participants reported that social-overload stress was reduced.