AIAny
Icon for item

NCP-ArchPreview Technical Report: Moving towards Latent Space Language Models through Next Concept Prediction

Adds a Next Concept Prediction objective that predicts discrete, product-quantized latent concepts spanning multiple tokens alongside next-token prediction, yielding faster pretraining convergence and measurable downstream gains with an 8.9B-parameter latent-space LLM.

Introduction

Most autoregressive LLM pretraining focuses only on next-token prediction. NCP-ArchPreview's core insight is to force the model to also predict higher-level discrete "concepts"—compact latent units that summarize several tokens—so the model learns semantic continuations at a coarser timescale and uses those predictions to guide token-level generation.

Key Findings
  • Joint NTP (next-token) + NCP (next-concept) speeds effective pretraining: Stage 1 reaches the final training loss of an OLMo-3-7B baseline after consuming only 51.3% of the token budget, implying ~1.95× token-budget convergence.
  • Strong downstream gains after full pretraining: +2.45 points macro-average vs OLMo-3-7B, including +5.99 on GSM8K and +4.28 on HumanEval, indicating improved reasoning and code capabilities.
  • Architecture & efficiency: scaled to ~8.94B parameters, trained on 5.73T Dolma-3 tokens; using ~85% of standard computation approaches the loss of a parameter-aligned 8.9B baseline.
  • Latent-space utility after pretraining: concept codebooks are small (~17M parameters) and can be updated for lightweight domain adaptation; injecting concept representations into a DFlash2 drafter increased mean accepted length by 4.17% with negligible overhead.
What Sets It Apart
  • Explicit concept-level supervision: constructs concept vectors by mean-pooling groups of four token encoder states and predicts future concepts at one-quarter the token rate, creating a concrete training signal above token-level statistics.
  • Product-quantized concept vocabulary: 32 codebooks provide a discrete yet differentiable concept space; predictions are differentiable weighted combinations of codewords, enabling end-to-end learning with standard autoregressive decoding.
  • Modular design for adaptation: token backbone (encoder/decoder) remains usable while a small VQ module can be re-trained for domain shifts, enabling cheaper adaptation than full model fine-tuning.
Who it's for — tradeoffs

Great fit if you want to reduce pretraining token/computation budgets while improving reasoning/code benchmarks, or if you need a small, updatable interface (the ≈17M-parameter VQ module) for domain adaptation. Look elsewhere if you require strictly minimal system complexity (adds a concept module, codebooks, and altered training pipeline) or if your deployment stack cannot accommodate the added latent-interface or slightly different generation pathway.

Method details
  • Pipeline: 16-layer token encoder → mean-pool every 4 token states into continuous concept vectors → 8-layer Concept Module predicts next concept → product quantization (32 codebooks) produces discrete concept vocab → 16-layer token decoder conditions on causally aligned concept predictions to produce tokens.
  • Reported metrics: Stage 1 Overall AVG 49.04 vs 46.59 for OLMo-3-7B; MMLU 64.80 (+2.58); GSM8K 45.26 (+5.99); HumanEval 31.38 (+4.28). Context length used in released checkpoints: 8192 tokens.

Information

  • Websitearxiv.org
  • OrganizationsShanghai AI Lab, LUMIA Lab, Shanghai Jiao Tong University
  • AuthorsJiaqi Cao, Chiyu Chen, Shuang Cheng, Xu Cheng, Beiya Dai, Yufan Feng, Kewen Ge, Ruijun Ge, Jiayi Huang, Yang Jiao
  • Published date2026/09/09

More Items

Progressively prunes and distills audio encoders for speech LLMs to cut inference cost while preserving decoder-facing embeddings, using behavioral probes, representation alignment, cross-scale distillation and LoRA finetuning; reports reduced macro-error on Chinese–English benchmarks.

Provides a diagnostic benchmark of 4,084 multiple-choice questions derived from longitudinal wearable time series, blood biomarkers, and demographics of 200 users to evaluate LLM reasoning on real-world wearable data. Uses 16 question types across complementary axes and a dual-grounding framework combining literature findings with population statistics.

Evaluates a weak teacher's RL-induced policy shift on the student's own rollouts and amplifies verifier-supported updates so stronger models can learn from weaker supervisors and surpass them. It rescales only verifier-supported policy-gradient components to preserve optimization fixed points while accelerating learning, reducing student updates versus standard RL or distillation.