REVIEW 27 cited by
Think before you speak: Training Language Models With Pause Tokens
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
Think before you speak: Training Language Models With Pause Tokens
read the original abstract
Language models generate responses by producing a series of tokens in immediate succession: the $(K+1)^{th}$ token is an outcome of manipulating $K$ hidden vectors per layer, one vector per preceding token. What if instead we were to let the model manipulate say, $K+10$ hidden vectors, before it outputs the $(K+1)^{th}$ token? We operationalize this idea by performing training and inference on language models with a (learnable) $\textit{pause}$ token, a sequence of which is appended to the input prefix. We then delay extracting the model's outputs until the last pause token is seen, thereby allowing the model to process extra computation before committing to an answer. We empirically evaluate $\textit{pause-training}$ on decoder-only models of 1B and 130M parameters with causal pretraining on C4, and on downstream tasks covering reasoning, question-answering, general understanding and fact recall. Our main finding is that inference-time delays show gains when the model is both pre-trained and finetuned with delays. For the 1B model, we witness gains on 8 of 9 tasks, most prominently, a gain of $18\%$ EM score on the QA task of SQuAD, $8\%$ on CommonSenseQA and $1\%$ accuracy on the reasoning task of GSM8k. Our work raises a range of conceptual and practical future research questions on making delayed next-token prediction a widely applicable new paradigm.
Forward citations
Cited by 27 Pith papers
-
Transformers Provably Learn to Internalize Chain-of-Thought
L-layer transformers under Log-ICoT curriculum provably learn k-parity with poly(n) samples and log k stages, matching explicit CoT efficiency without inference overhead.
-
Does Verbose Chain-of-Thought Really Help? In-Distribution Evidence that Content, Not Length, Matters
In-distribution sampling across 25 models and controlled interventions with DAG-verified content show that semantic reasoning and validation content, not token count, drive CoT gains.
-
Demystifying Hidden-State Recurrence: Switchable Latent Reasoning with On-Policy Reinforcement Learning
SWITCH uses explicit <swi> and </swi> boundary tokens to make latent chain-of-thought compatible with on-policy RL (GRPO) and open to causal mechanistic probing, outperforming prior hidden-state recurrence methods.
-
Forecasting Future Behavior as a Learning Task
Behavior Forecasters trained on LRM trajectories outperform larger models in predicting repeatability and input sensitivity at low cost.
-
LatentOmni: Rethinking Omni-Modal Understanding via Unified Audio-Visual Latent Reasoning
LatentOmni proposes a latent-space cross-modal reasoning framework that uses feature-level supervision and Omni-Sync Position Embedding to align and synchronize audio-visual latents, supported by a new 35K interleaved...
-
Dynamic Latent Routing
Dynamic Latent Routing jointly learns discrete latent codes, routing policies, and model parameters via dynamic search to match or exceed supervised fine-tuning by 6.6 points on average in low-data settings across fou...
-
4DThinker: Thinking with 4D Imagery for Dynamic Spatial Understanding
4DThinker enables VLMs to perform dynamic spatial reasoning by thinking with 4D latent mental imagery using new fine-tuning and reinforcement learning methods.
-
Latent Abstraction for Retrieval-Augmented Generation
LAnR unifies retrieval-augmented generation inside a single LLM by deriving dense retrieval vectors from a [PRED] token's hidden states and using entropy to adaptively stop retrieval, outperforming prior RAG on six QA...
-
AI Achieves a Perfect LSAT Score
Language models achieve a perfect LSAT score, with experiments showing that internal thinking phases and a fine-tuned process reward model are key to high performance on logical reasoning questions.
-
Training Large Language Models to Reason in a Continuous Latent Space
Coconut lets LLMs perform reasoning directly in continuous latent space by recycling hidden states as inputs, outperforming standard chain-of-thought on search-intensive logical tasks with better accuracy-efficiency t...
-
Why Limit the Residual Stream to Layers and Not Tokens? Persistent Memory for Continuous Latent Reasoning
AGCLR extends CoCoNuT with a gated concept stream for persistent memory to fix fact loss in latent reasoning, yielding improvements on reasoning benchmarks as depth increases.
-
Test-Time Compute Scaling for ASR with Depth-Conditioned Looped Transformers
LARM enables test-time compute scaling in non-autoregressive ASR via depth-conditioned looping with CTC checkpoints, supervision embeddings, FiLM conditioning, and posterior feedback, yielding lower WER on LibriSpeech...
-
TTE-Flash: Accelerating Reasoning-based Multimodal Representations via Think-Then-Embed Tokens
TTE-Flash trains latent think tokens with CoT generation loss and embedding tokens with contrastive loss to deliver high-performance multimodal representations without generating explicit reasoning at inference time.
-
4DThinker: Thinking with 4D Imagery for Dynamic Spatial Understanding
4DThinker enables VLMs to perform dynamic spatial reasoning by internally simulating 4D imagery in latent space, outperforming prior text-based and modular approaches.
-
HypEHR: Hyperbolic Modeling of Electronic Health Records for Efficient Question Answering
HypEHR is a hyperbolic embedding model for EHR data that uses Lorentzian geometry and hierarchy-aware pretraining to answer clinical questions nearly as well as large language models but with much smaller size.
-
SeLaR: Selective Latent Reasoning in Large Language Models
SeLaR selectively applies latent soft reasoning in LLMs via entropy gating and contrastive regularization, outperforming standard CoT on five benchmarks without training.
-
Representational Collapse in Multi-Agent LLM Committees: Measurement and Diversity-Aware Consensus
LLM agent committees exhibit representational collapse with mean cosine similarity of 0.888, and diversity-aware consensus reaches 87% accuracy on GSM8K versus 84% for self-consistency at lower cost.
-
Next-Latent Prediction Transformers Learn Compact World Models
NextLat augments next-token prediction with latent next-state prediction, theoretically converging latents to belief states and showing empirical gains in world modeling, reasoning, planning, and faster inference via ...
-
Can Aha Moments Be Fake? Towards Quantifying Decorative and True Thinking in Chain-of-Thought
LLMs interleave true causal reasoning steps with decorative ones in CoT, with only ~2.3% of steps having high causal impact on AIME for Qwen-2.5, and a steering direction can force internal use of specific steps.
-
Mind-Paced Speaking: A Dual-Brain Approach to Real-Time Reasoning in Spoken Language Models
MPS proposes a dual-brain architecture separating formulation reasoning from articulation to achieve real-time CoT in SLMs with accuracy comparable to full pre-computation but much lower latency.
-
Compressed Chain of Thought: Efficient Reasoning Through Dense Representations
CCoT generates variable-length continuous contemplation tokens that compress explicit reasoning chains, enabling additional dense reasoning and accuracy gains in off-the-shelf language models while allowing adaptive c...
-
MeTHanol: Modularized Thinking Language Models with Intermediate Layer Thinking, Decoding and Bootstrapping Reasoning
MeTHanol fine-tunes an intermediate LLM layer to generate thoughts in a first pass, then uses those thoughts for a second-pass answer, showing gains on Theory of Mind and vignette tasks plus adaptation to character prompts.
-
Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters
An adaptive compute-optimal strategy for scaling LLM test-time compute achieves over 4x efficiency gains versus best-of-N and lets smaller models outperform 14x larger ones on some problems.
-
MIRAGE: Mobile Agents with Implicit Reasoning and Generative World Models
MIRAGE compresses explicit chain-of-thought into latent vectors and adds a generative world model to predict future interface states, matching explicit reasoning performance with 3-5x fewer tokens on Android benchmarks.
-
NoisyCoconut: Counterfactual Consensus via Latent Space Reasoning
Injecting noise into LLM latent trajectories creates diverse reasoning paths whose agreement acts as a confidence signal for selective abstention, cutting error rates from 40-70% to under 15% on math tasks.
-
HALO: Hybrid Adaptive Latent Reasoning for Language Models
Selective second-stage latent refinement on budgeted scored tokens beats uniform fixed-1 and fixed-2 refinement on frozen Phi-4-mini for average MMLU-Pro/GPQA score at lower average applied refine steps.
-
ConFu: Contemplate the Future for Better Speculative Sampling
ConFu boosts speculative decoding acceptance rates 8-20% over EAGLE-3 by letting draft models use contemplate tokens and MoE to anticipate future generation direction.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.