Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution.
For example: A model processes several pause tokens before answering a math question. Those extra positions let it perform additional computation, even though they do not display a worked solution.
Pause tokens add computation before the answer without choosing a different plan token for each step. Planning tokens instead select a learned token before generating each written reasoning step. The latter still produces a textual chain of thought. In the pause-token experiments, the training setup matters: gains were strongest when pauses were used in both pretraining and fine-tuning, not simply inserted into an unchanged model.
Sources: Pause-token paper · Planning-token paper, §2
Paper details: Think before you speak: Training language models with pause tokens · Guiding Language Model Reasoning with Planning Tokens
Dates usually refer to the first arXiv submission. Search or filter these papers →
Gives a recurrent language model extra hidden-state updates by inserting special tokens after each word. Tests whether these updates improve next-word prediction.
Date source: AITP presentation date; the earlier public release date of the conference abstract is unverified. Linked arXiv version: May 14, 2024.
Introduces pause tokens that give a language model additional computational positions before it must produce an answer.
Adds a learned planning token before each written reasoning step, then uses that token to guide the step.
Trains on ordinary text by generating written rationales before predicting what comes next. Learns from how much those rationales improve subsequent token predictions.
Studies whether meaningless filler-token positions can provide transformers with additional hidden computation without carrying readable reasoning content.
Interleaves latent and textual tokens so a model can combine compressed internal computation with readable reasoning steps.
Adds hidden decoding tokens during pretraining, then discards their key/value cache entries after use. Keeps the original tokens' entries to preserve longer-range context.
Adds trainable dummy tokens to a pretrained transformer so attention can use them for extra computation during decoding. The tokens can be inserted in different positions at inference time.
Uses latent visual tokens as an internal imagery channel for multimodal reasoning without rendering every intermediate state.
Analyzes how skipping intermediate steps weakens learning signals in a simplified parity-task model. Introduces NatBool-DAG to test logical dependencies and ALiCoT, which supervises learned token states with intermediate logical labels or explicit-CoT representations.
Supervises intermediate thinking states that bridge textual reasoning traces and latent computation.
Routes intermediate computation between latent and discrete representations to trade off reasoning quality, readability, and token cost.
Moves pause-token computation into a parallel, weight-shared prediction stream, providing additional internal compute without increasing generated sequence length or key/value-cache length.
Places pause tokens at reasoning-step boundaries and masks their training loss. Studies how this affects learning new tasks while retaining earlier capabilities, including during reinforcement learning.
Inserts a hidden-state refinement step between consecutive text tokens in a recurrent transformer. Compares this extra thinking time with adding physical layers or transformer loops under matched computation budgets.
LIRSeg generates a fixed sequence of learned special tokens before a target description and bounding box used for segmentation. Spatial alignment grounds the tokens' hidden states in relevant image regions, followed by training with segmentation rewards.
Analyzes tasks for which a fixed-depth transformer needs an informative chain of thought to exceed a task-specific accuracy limit. Combines circuit-based results with experiments that replace, erase, or shuffle reasoning tokens, including comparisons with filler tokens.
Compresses each completed block of generated reasoning into the key/value cache of learned pause tokens, then discards the block's original cache. Later tokens attend to the compact replacement.