Skip to content
conftrace_
2026 EMNLP EMNLP 2026

Trajectory Self-Feedback in Diffusion Language Models: A Systematic Survey

AbstractDiffusion language models generate text through iterative denoising, allowing intermediate predictions to influence later denoising steps, a recurring practice we call trajectory self-feedback (TSF). TSF is adopted under inconsistent terminology, evaluated on incompatible benchmarks, and studied by disconnected communities, yet no prior review characterises it reproducibly. We present a systematic literature review under an evidence-based methodology: six pre-specified questions, a documented six-library search with snowballing, explicit selection and quality criteria, and a structured synthesis of 24 primary studies across seven clusters, with a released audit trail. We find: (i) a model-agnostic taxonomy resolves four operator families: Prediction-Space Concatenation, Confidence-weighted Path Signals, Reward-Reinforced Conditioning, and Correction-Targeted Feedback, that the literature conflates as self-conditioning; (ii) evidence of benefit is real but fragmented across incomparable settings; (iii) within our search window, no study jointly varies the TSF mechanism and measures faithfulness on knowledge-intensive tasks; and (iv) a discrete/continuous asymmetry consistent with the sampling-wall hypothesis is visible but never tested. We conclude with eight open problems, a dependency map, and review-derived diagnostics (Commitment Index, Correction Capacity) and a Standardised Evaluation Protocol (SEP-DLM) to make TSF measurable across quality, correction, trajectory commitment, and faithfulness.

More from EMNLP 2026

All EMNLP 2026 papers →