2nd/10_Wiki/Topics/Encoder-Decoder-Inconsistency.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-encoder-decoder-inconsistency	Encoder Decoder Inconsistency	10_Wiki/Topics	verified	self	encoder-decoder-mismatch seq2seq-inconsistency	none	A	0.9	applied	nlp transformer training decoding		2026-05-10	pending	language framework python pytorch

Encoder Decoder Inconsistency

매 한 줄

"매 encoder가 본 분포 ≠ decoder가 생성하는 분포". Seq2seq training 시 encoder는 ground-truth context를 보지만 decoder는 inference에서 자기 prediction을 다시 입력으로 받기 때문에 train/inference 간 distribution shift가 발생한다. 매 exposure bias 의 근본 원인.

매 핵심

매 정의

• Train: decoder input = teacher-forced ground truth.
• Inference: decoder input = previously generated token.
• Gap: 매 error compounds along sequence — early mistake → later tokens conditioned on out-of-distribution prefix.

매 표현

• Exposure bias (Ranzato 2016).
• Schedule sampling 의 motivation.
• Hallucination 의 한 원인 (특히 long-form generation).

매 응용

1. NMT (Neural Machine Translation) — 매 long sentence translation degradation.
2. Summarization — repetition / drift.
3. Speech recognition — RNN-T vs CTC trade-off.
4. Code generation — 매 long completion 의 syntax break.

💻 패턴

Scheduled Sampling

import torch
import torch.nn.functional as F

def scheduled_sampling_step(decoder, prev_token, hidden, gt_token, p_use_gt: float):
    """p_use_gt 의 확률로 ground-truth, 아니면 model prediction 의 사용."""
    if torch.rand(1).item() < p_use_gt:
        input_tok = gt_token
    else:
        with torch.no_grad():
            logits, _ = decoder(prev_token, hidden)
            input_tok = logits.argmax(dim=-1)
    out_logits, hidden = decoder(input_tok, hidden)
    return out_logits, hidden

Minimum Risk Training

def mrt_loss(model, src, refs, n_samples=8):
    """매 sequence-level loss 의 — 매 sampled hypotheses 에 대해 risk minimize."""
    hyps = [model.sample(src) for _ in range(n_samples)]
    risks = torch.tensor([1 - bleu(h, refs) for h in hyps])
    log_probs = torch.stack([model.log_prob(h, src) for h in hyps])
    weights = F.softmax(log_probs, dim=0)
    return (weights * risks).sum()

Self-distillation Fix

def self_distill(student, teacher, src, T=2.0):
    """매 teacher 가 자기 생성한 sequence 의 사용 — 매 train/inference gap 축소."""
    with torch.no_grad():
        gen = teacher.generate(src, do_sample=True, top_p=0.9)
        teacher_logits = teacher(src, gen).logits
    student_logits = student(src, gen).logits
    return F.kl_div(
        F.log_softmax(student_logits / T, dim=-1),
        F.softmax(teacher_logits / T, dim=-1),
        reduction="batchmean",
    ) * T * T

Beam Search with Length Penalty

def length_penalty(score, length, alpha=0.7):
    """GNMT length penalty — 매 short hypothesis 의 bias 보정."""
    return score / ((5 + length) ** alpha / (5 + 1) ** alpha)

Contrastive Decoding

def contrastive_decode(big, small, prompt, alpha=0.5):
    """매 large model logit − small model logit — 매 expert/amateur gap 의 강조."""
    big_logits = big(prompt).logits[:, -1]
    small_logits = small(prompt).logits[:, -1]
    return big_logits - alpha * small_logits

매 결정 기준

상황	Approach
Short sequence (<32)	Teacher forcing 충분
Long sequence	Scheduled sampling / MRT
Production NMT	Beam + length penalty + coverage
LLM long-form	Contrastive decoding / self-distillation

기본값: teacher forcing + 1k step warmup 이후 scheduled sampling.

🔗 Graph

• 부모: Sequence-to-Sequence · Transformer
• 변형: Scheduled-Sampling · Minimum-Risk-Training
• 응용: Neural-Machine-Translation · Abstractive-Summarization
• Adjacent: Exposure-Bias · Hallucination · Beam-Search

🤖 LLM 활용

언제: long-form generation 의 quality issue 분석 시. Train/eval BLEU gap 의 진단. 언제 X: 매 short classification — 매 inconsistency 의 무관.

❌ 안티패턴

• Pure teacher forcing forever: 매 inference distribution 의 미본 채 deploy.
• Greedy decoding only: 매 early mistake 의 lock-in.
• No length normalization: beam 의 short hypothesis bias.

🧪 검증 / 중복

• Verified (Ranzato et al. 2016, Bengio et al. 2015 scheduled sampling).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — encoder/decoder distribution shift + scheduled sampling/MRT/contrastive decoding