TokAlign++ learns token alignments between LLM vocabularies from monolingual representations to enable faster adaptation, better text compression, and effective token-level distillation across 15 languages with minimal steps.
Title resolution pending
26 Pith papers cite this work. Polarity classification is still indexing.
26
Pith papers citing it
citation-role summary
background 2
citation-polarity summary
roles
background 1polarities
background 1representative citing papers
Introduces TBPO, which derives a Bregman-divergence density-ratio matching objective for token-level preference optimization that generalizes DPO while preserving the induced optimal policy.
BWLA is the first post-training quantization method for LLMs that achieves 1-bit weights paired with low-bit activations such as 6 bits, using OKT to reshape weights and suppress activation tails plus PSP for low-rank refinement.
Uncertainty and correctness in LLMs are encoded by distinct feature populations, with suppression of confounded features improving accuracy and reducing entropy.
LoRA adapters should be scaled by 1/sqrt(rank) rather than 1/rank to stabilize learning and enable effective use of higher ranks during fine-tuning of large language models.
Pruning MoE LLMs according to the stationary distribution of a Markov chain over (layer, expert) routing transitions retains more task performance than local importance heuristics, with up to ~3.5% relative gains over the strongest baseline at 50% compression.
SDM sparsifies the Gated DeltaNet update rule to enable 1000x larger recurrent memory states at iso-FLOP, improving long-context recall and short-context reasoning over GDN and matching full attention at 8B scale.
A two-stream Transformer variant that separates state storage from next-token prediction improves validation loss and downstream task performance by 2-3 points over standard Transformers.
With enough compute, large models benefit from training on unfiltered data that includes low-quality and distractor examples instead of requiring high-quality filtered data.
GTLM injects graph-aware attention biases into LLMs using only 0.015% extra parameters, enabling native graph processing that matches 7B models with a 1B model on text-attributed graph benchmarks.
TPAW uses teams of current and historical model checkpoints that collaborate and compete, plus adaptive weightings for responses and players, to improve self-supervised LLM alignment and outperform baselines.
LLMs contain identifiable COCO neurons that enable implicit self-correction against stereotypes; targeted editing of these neurons improves fairness and robustness to jailbreaks while preserving generation quality.
Decision theory shows that LLM cascades are structurally limited by always incurring the cheap model's cost before deciding to escalate, with the best performance given by the envelope of pairwise cascades rather than fixed chains or many stages.
ZAYA1-8B is a reasoning MoE model with 700M active parameters that matches larger models on math and coding benchmarks and reaches 91.9% on AIME'25 via Markovian RSA test-time compute.
InfoLaw models pretraining as information accumulation where quality sets information density and repetition causes scale-dependent diminishing returns, predicting loss with low error on unseen mixtures and larger scales up to 7B models and 425B tokens.
GRASPrune removes 50% of parameters from LLaMA-2-7B via global gating and projected straight-through estimation, reaching 12.18 WikiText-2 perplexity and competitive zero-shot accuracy after four epochs on 512 calibration sequences.
DPOP is a new loss function that prevents DPO from lowering preferred response likelihoods and outperforms standard DPO on diverse datasets, MT-Bench, and enables Smaug-72B to exceed 80% on the Open LLM Leaderboard.
Large reasoning models show a negative correlation between token entropy and logit-gradient norms, and the paper uses this 'Entropy-Gradient Inversion' as a reward regularizer (CorR-PO) in RL training.
Byte modeling incurs greater scaling overhead for masked diffusion than autoregressive models because the diffusion objective destroys local byte contiguity needed to resolve semantics.
Continual pre-training on a German medical corpus lets 7B models close much of the performance gap with 24B general models on medical benchmarks, though merging introduces some language mixing and verbosity.
MASS-RAG uses distinct agents for evidence summarization, extraction, and reasoning, then synthesizes their outputs to improve answer quality over standard RAG baselines on four benchmarks, especially when evidence is distributed.
SmolLM2 is a 1.7B-parameter language model that outperforms Qwen2.5-1.5B and Llama3.2-1B after overtraining on 11 trillion tokens using custom FineMath, Stack-Edu, and SmolTalk datasets in a multi-stage pipeline.
Fisher information from the target data distribution supplies a task-dependent criterion for selecting LoRA directions that outperforms weight-magnitude heuristics.
Multilingual pooling for quality classifiers outperforms monolingual baselines in rank stability and accuracy for LLM pretraining data selection across high- and low-resource languages.
citing papers explorer
-
TokAlign++: Advancing Vocabulary Adaptation via Better Token Alignment
TokAlign++ learns token alignments between LLM vocabularies from monolingual representations to enable faster adaptation, better text compression, and effective token-level distillation across 15 languages with minimal steps.
-
TokenRatio: Principled Token-Level Preference Optimization via Ratio Matching
Introduces TBPO, which derives a Bregman-divergence density-ratio matching objective for token-level preference optimization that generalizes DPO while preserving the induced optimal policy.
-
BWLA: Breaking the Barrier of W1AX Post-Training Quantization for LLMs
BWLA is the first post-training quantization method for LLMs that achieves 1-bit weights paired with low-bit activations such as 6 bits, using OKT to reshape weights and suppress activation tails plus PSP for low-rank refinement.
-
Are LLM Uncertainty and Correctness Encoded by the Same Features? A Functional Dissociation via Sparse Autoencoders
Uncertainty and correctness in LLMs are encoded by distinct feature populations, with suppression of confounded features improving accuracy and reducing entropy.
-
A Rank Stabilization Scaling Factor for Fine-Tuning with LoRA
LoRA adapters should be scaled by 1/sqrt(rank) rather than 1/rank to stabilize learning and enable effective use of higher ranks during fine-tuning of large language models.
-
It Takes a MAESTRO To Prune Bad Experts
Pruning MoE LLMs according to the stationary distribution of a Markov chain over (layer, expert) routing transitions retains more task performance than local importance heuristics, with up to ~3.5% relative gains over the strongest baseline at 50% compression.
-
Sparse Delta Memory: Scaling the State of Linear RNNs through Sparsity
SDM sparsifies the Gated DeltaNet update rule to enable 1000x larger recurrent memory states at iso-FLOP, improving long-context recall and short-context reasoning over GDN and matching full attention at 8B scale.
-
The State-Prediction Separation Hypothesis
A two-stream Transformer variant that separates state storage from next-token prediction improves validation loss and downstream task performance by 2-3 points over standard Transformers.
-
A Bitter Lesson for Data Filtering
With enough compute, large models benefit from training on unfiltered data that includes low-quality and distractor examples instead of requiring high-quality filtered data.
-
Teaching LLMs to See Graphs: Unifying Text and Structural Reasoning
GTLM injects graph-aware attention biases into LLMs using only 0.015% extra parameters, enabling native graph processing that matches 7B models with a 1B model on text-attributed graph benchmarks.
-
Team-Based Self-Play With Dual Adaptive Weighting for Fine-Tuning LLMs
TPAW uses teams of current and historical model checkpoints that collaborate and compete, plus adaptive weightings for responses and players, to improve self-supervised LLM alignment and outperform baselines.
-
Modeling Implicit Conflict Monitoring Mechanisms against Stereotypes in LLMs
LLMs contain identifiable COCO neurons that enable implicit self-correction against stereotypes; targeted editing of these neurons improves fairness and robustness to jailbreaks while preserving generation quality.
-
Is Escalation Worth It? A Decision-Theoretic Characterization of LLM Cascades
Decision theory shows that LLM cascades are structurally limited by always incurring the cheap model's cost before deciding to escalate, with the best performance given by the envelope of pairwise cascades rather than fixed chains or many stages.
-
ZAYA1-8B Technical Report
ZAYA1-8B is a reasoning MoE model with 700M active parameters that matches larger models on math and coding benchmarks and reaches 91.9% on AIME'25 via Markovian RSA test-time compute.
-
InfoLaw: Information Scaling Laws for Large Language Models with Quality-Weighted Mixture Data and Repetition
InfoLaw models pretraining as information accumulation where quality sets information density and repetition causes scale-dependent diminishing returns, predicting loss with low error on unseen mixtures and larger scales up to 7B models and 425B tokens.
-
GRASPrune: Global Gating for Budgeted Structured Pruning of Large Language Models
GRASPrune removes 50% of parameters from LLaMA-2-7B via global gating and projected straight-through estimation, reaching 12.18 WikiText-2 perplexity and competitive zero-shot accuracy after four epochs on 512 calibration sequences.
-
Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive
DPOP is a new loss function that prevents DPO from lowering preferred response likelihoods and outperforms standard DPO on diverse datasets, MT-Bench, and enables Smaug-72B to exceed 80% on the Open LLM Leaderboard.
-
Entropy-Gradient Inversion: Moving Toward Internal Mechanism of Large Reasoning Models
Large reasoning models show a negative correlation between token entropy and logit-gradient norms, and the paper uses this 'Entropy-Gradient Inversion' as a reward regularizer (CorR-PO) in RL training.
-
The Efficiency Gap in Byte Modeling
Byte modeling incurs greater scaling overhead for masked diffusion than autoregressive models because the diffusion objective destroys local byte contiguity needed to resolve semantics.
-
Can Continual Pre-training Bridge the Performance Gap between General-purpose and Specialized Language Models in the Medical Domain?
Continual pre-training on a German medical corpus lets 7B models close much of the performance gap with 24B general models on medical benchmarks, though merging introduces some language mixing and verbosity.
-
MASS-RAG: Multi-Agent Synthesis Retrieval-Augmented Generation
MASS-RAG uses distinct agents for evidence summarization, extraction, and reasoning, then synthesizes their outputs to improve answer quality over standard RAG baselines on four benchmarks, especially when evidence is distributed.
-
SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model
SmolLM2 is a 1.7B-parameter language model that outperforms Qwen2.5-1.5B and Llama3.2-1B after overtraining on 11 trillion tokens using custom FineMath, Stack-Edu, and SmolTalk datasets in a multi-stage pipeline.
-
Learning in the Fisher Subspace: A Guided Initialization for LoRA Fine-Tuning
Fisher information from the target data distribution supplies a task-dependent criterion for selecting LoRA directions that outperforms weight-magnitude heuristics.
-
Toward Cross-Lingual Quality Classifiers for Multilingual Pretraining Data Selection
Multilingual pooling for quality classifiers outperforms monolingual baselines in rank stability and accuracy for LLM pretraining data selection across high- and low-resource languages.
-
Gemma: Open Models Based on Gemini Research and Technology
Gemma introduces open 2B and 7B LLMs derived from Gemini technology that beat comparable open models on 11 of 18 text tasks and come with safety assessments.
-
Gemma 2: Improving Open Language Models at a Practical Size
Gemma 2 models achieve leading performance at their sizes by combining established Transformer modifications with knowledge distillation for the 2B and 9B variants.