# Parallel Decoder Transformer: Planner-Conditioned Latent Coordination for Model-Intrinsic Parallel Generation

arXiv: https://arxiv.org/abs/2512.10054 · HTML: https://mlsanity.com/paper/2512.10054/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 13 of its 19 references that are in the map.
- First posted on arXiv 10 Dec 2025.

## Builds on (most cited first, 13 of 13)

- [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://mlsanity.com/paper/2005.11401/) (2020)
- [Neural Discrete Representation Learning](https://mlsanity.com/paper/1711.00937/) (2017)
- [Fast Inference from Transformers via Speculative Decoding](https://mlsanity.com/paper/2211.17192/) (2022)
- [Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads](https://mlsanity.com/paper/2401.10774/) (2024)
- [Sequence-Level Knowledge Distillation](https://mlsanity.com/paper/1606.07947/) (2016)
- [EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty](https://mlsanity.com/paper/2401.15077/) (2024)
- [Non-Autoregressive Neural Machine Translation](https://mlsanity.com/paper/1711.02281/) (2017)
- [Better & Faster Large Language Models via Multi-token Prediction](https://mlsanity.com/paper/2404.19737/) (2024)
- [Skeleton-of-Thought: Prompting LLMs for Efficient Parallel Generation](https://mlsanity.com/paper/2307.15337/) (2023)
- [Hogwild! Inference: Parallel LLM Generation via Concurrent Attention](https://mlsanity.com/paper/2504.06261/) (2025)
- [Learning to Keep a Promise: Scaling Language Model Decoding Parallelism with Learned Asynchronous Decoding](https://mlsanity.com/paper/2502.11517/) (2025)
- [APAR: LLMs Can Do Auto-Parallel Auto-Regressive Decoding](https://mlsanity.com/paper/2401.06761/) (2024)
- [Group Think: Multiple Concurrent Reasoning Agents Collaborating at Token Level Granularity](https://mlsanity.com/paper/2505.11107/) (2025)

## Abstract (arXiv)

Autoregressive language models expose one causal token frontier, even when the requested document contains sections that could be developed concurrently. Existing parallel-generation systems arrange external branches around an otherwise unchanged model. We instead formulate model-intrinsic parallel generation: a single trained architecture owns multiple causal frontiers and produces one next-token distribution for each frontier in every synchronized decoding round. The Parallel Decoder Transformer (PDT) retains a frozen shared lower knowledge trunk and replaces the upper trunk with three independently parameterized physical decoder stacks. A prompt-time set planner produces three unordered continuous outlines, each hard-routed to one decoder as persistent Plan-KV memory, while a finite product-quantized notes bus carries block-delayed latent messages among the decoders. Autoregression is preserved within each lane; same-round lane tokens are conditionally independent given the source, plans, private histories, and previously committed messages. We specify source-grounded supervision for long-form historical exposition with single-owner cited facts and token-aligned cross-lane dependencies, a composite objective, a staged curriculum, and preregistered causal evaluations: plan swap and removal, delayed-message ablation, a parameter-matched self-only control, dependency-token likelihood, and blinded human fact audits. The architecture and evaluation pipeline are implemented; scientific training and held-out evaluation are in progress. This paper presents the theory, design, and falsifiable protocol, not a positive empirical result.