AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Beyond Multi-Token Prediction: Pretraining LLMs with Future Summaries

Divyat Mahajan, Sachin Goyal, Badr Youbi Idrissi, Mohammad Pezeshki, Ioannis Mitliagkas, David Lopez-Paz, Kartik Ahuja

ICLR 2026 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Use learned reverse-LM future summary instead of next-token prediction onlyFSP-RevLM · ARC-Easy (8B)pass@1: 71.8% -> 76.6%Improved
Use FSP-RevLM instead of NTPFSP-RevLM · GSM8K (8B)pass@16: 0.716 -> 0.705Hurt
Use FSP-RevLM instead of MTPFSP-RevLM · MATH (8B)pass@16: 0.309 -> 0.351Improved
Use FSP-RevLM instead of DeepSeek-MTP at small scaleFSP-RevLM · ARC and code tasks (3B)pass@1/pass@16: HumanEval+ 0.348 -> 0.305Hurt
Predict random future tokens (MTP-Skip) instead of the immediate next token (ablation)MTP · GSM8K (8B)pass@16: 0.678 -> 0.598 (tau 32)Hurt
Use bag-of-words future summary instead of immediate-token MTP target (ablation)MTP · GSM8K (8B)pass@16: 0.678 -> 0.714 (tau 100)Improved
Lengthen the BoW window from 12 to 100 (ablation)FSP-BoW · ARC-Challenge (8B)pass@1: 0.562 -> 0.459Hurt
Use a bag-of-words summary on path-star graphs instead of MTP (ablation)MTP · Path-star G(2,8)exact path accuracy: 0.48 -> 1.00Improved
Add more auxiliary heads to MTP (ablation)MTP · Path-star G(2,8)exact path accuracy: 0.48 (4 heads: 0.48)No clear effect
Use RevLM summary instead of BoW on sibling discovery with many components (ablation)FSP-BoW · Sibling discoverysteps to convergence vs NTPImproved
Remove tf-idf reweighting from FSP-BoW (ablation)FSP-BoW tau 12 · GSM8K (3B)pass@16: 0.419 -> 0.431Mixed
Use a shallower reverse-LM depth (depth 2) in FSP-RevLM (ablation)FSP-RevLM · 3B taskspass@16 / pass@1: MATH 0.212 -> 0.217; ARC-Easy 0.277 -> 0.265Mixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy