MultiHashFormer enables hash-based autoregression in LMs by encoding tokens as multi-hash signatures, outperforming standard Transformers at 100M-3B scales while keeping parameter count constant for multilingual expansion.
Canonical reference
HellaSwag: Can a Machine Really Finish Your Sentence? , booktitle =
Canonical reference. 71% of citing Pith papers cite this work as background.
95
Pith papers citing it
453
external citations · Crossref
Background
71% of classified citations
citation-role summary
background 5
dataset 2
citation-polarity summary
co-cited works
representative citing papers
Output vector editing on MLP neurons suppresses memorization in LLMs up to 87.9% on 6831 sequences in OLMo-7B with a 2.7x gap over zero ablation, ensemble covering 96.5%.
Repetition rate mismatch between small-scale proxies and target budgets is the main reason data mixture experiments do not scale; a subsampling procedure that equalizes repetition rates recovers optimal mixtures from 1/16-scale experiments.
The paper introduces an outlier-injection attack that induces targeted weight collapse in LLMs under advanced quantization schemes including AWQ, GPTQ, and GGUF I-quants.
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.
ConQuR is a post-training rotation calibration technique that aligns activations to hypercube corners via Procrustes optimization and online updates, delivering competitive LLM quantization performance without end-to-end training or offline activation storage.
EdgeFlowerTune is a real-device benchmark that jointly assesses model quality and system costs for federated LLM fine-tuning on edge hardware using three protocols: Quality-under-Budget, Cost-to-Target, and Robustness.
Gaussian Kernel Attention replaces learned QKV projections with a Gaussian RBF kernel on per-head token features, using 0.42x parameters and 0.49x FLOPs while showing competitive language modeling performance at depth 20.
Preconditioned delta-rule models with a diagonal curvature approximation improve upon standard DeltaNet, GDN, and KDA by better approximating the test-time regression objective.
Structured rephrasing prompts (tables, math problems, FAQs, tutorials) beat curated web text and prior synthetic data, and small 1B generators match or beat 27B generators for pretraining-data synthesis.
ProofGrid is a new benchmark for LLM reasoning that uses machine-checkable proofs in minimal formal notation, revealing progress on basic tasks but major gaps in complex combinatorial and synthesis reasoning.
LLM unlearning is reframed as inadvertently installing backdoor triggers on forget-tokens; Random Noise Augmentation is introduced as a defense that improves robustness with theoretical guarantees.
DeepSeek-V2 delivers top-tier open-source LLM performance using only 21B active parameters by compressing the KV cache 93.3% and cutting training costs 42.5% via MLA and DeepSeekMoE.
RouterBench supplies a standardized benchmark, 405k+ inference dataset, theoretical framework, and comparative analysis for multi-LLM routing systems.
Iterative self-rewarding via LLM-as-Judge in DPO training on Llama 2 70B improves instruction following and self-evaluation, outperforming GPT-4 on AlpacaEval 2.0.
GAIA benchmark shows humans at 92% accuracy on simple real-world questions far outperform current AI systems at 15%, proposing this gap as a key milestone for general AI.
Chain-of-Thought reasoning in LLMs is often unfaithful, with models relying on it variably by task and less so as models scale larger.
OPT releases open decoder-only transformers up to 175B parameters that match GPT-3 performance at one-seventh the carbon cost, along with code and training logs.
A stateless, SVD-free regularizer approximates polar factors to induce low-rank weight structure during training, enabling better post-training compression of vision models and LLMs at under 8% overhead.
Optimal layer-patching order for boomerang distillation is a shortest path on a KL-weighted Boolean lattice; greedy KLPatch and simple sequential orders often yield near-optimal interpolations.
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.
FlexMoE produces nested pruned subnetworks for MoE LLMs across budgets via channel importance ranking and discrete action learning, plus one mid-budget recovery fine-tune, retaining 99.8% performance at 50% expert parameter pruning.
Aurora is a leverage-aware spectral optimizer that enforces uniform row norms in matrix updates while preserving Muon's polar geometry, outperforming Muon and achieving SOTA among spectral methods on modded-nanoGPT.
ConSA learns FA/SWA allocation via L0 masks and augmented Lagrangian constraints, outperforming rule-based baselines on 0.6B and 1.7B models with consistent layer patterns.
citing papers explorer
-
MultiHashFormer: Hash-based Generative Language Models
MultiHashFormer enables hash-based autoregression in LMs by encoding tokens as multi-hash signatures, outperforming standard Transformers at 100M-3B scales while keeping parameter count constant for multilingual expansion.
-
Output Vector Editing for Memorization Mitigation in Large Language Models
Output vector editing on MLP neurons suppresses memorization in LLMs up to 87.9% on 6831 sequences in OLMo-7B with a 2.7x gap over zero ablation, ensemble covering 96.5%.
-
Repetition Mismatch: Why Data Mixture Experiments Don't Scale and How to Fix Them
Repetition rate mismatch between small-scale proxies and target budgets is the main reason data mixture experiments do not scale; a subsampling procedure that equalizes repetition rates recovers optimal mixtures from 1/16-scale experiments.
-
Widening the Gap: Exploiting LLM Quantization via Outlier Injection
The paper introduces an outlier-injection attack that induces targeted weight collapse in LLMs under advanced quantization schemes including AWQ, GPTQ, and GGUF I-quants.
-
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.
-
ConQuR: Corner Aligned Activation Quantization via Optimized Rotations for LLMs
ConQuR is a post-training rotation calibration technique that aligns activations to hypercube corners via Procrustes optimization and online updates, delivering competitive LLM quantization performance without end-to-end training or offline activation storage.
-
EdgeFlowerTune: Evaluating Federated LLM Fine-Tuning Under Realistic Edge System Constraints
EdgeFlowerTune is a real-device benchmark that jointly assesses model quality and system costs for federated LLM fine-tuning on edge hardware using three protocols: Quality-under-Budget, Cost-to-Target, and Robustness.
-
Projection-Free Transformers via Gaussian Kernel Attention
Gaussian Kernel Attention replaces learned QKV projections with a Gaussian RBF kernel on per-head token features, using 0.42x parameters and 0.49x FLOPs while showing competitive language modeling performance at depth 20.
-
Preconditioned DeltaNet: Curvature-aware Sequence Modeling for Linear Recurrences
Preconditioned delta-rule models with a diagonal curvature approximation improve upon standard DeltaNet, GDN, and KDA by better approximating the test-time regression objective.
-
How Can We Synthesize High-Quality Pretraining Data? A Systematic Study of Prompt Design, Generator Model, and Source Data
Structured rephrasing prompts (tables, math problems, FAQs, tutorials) beat curated web text and prior synthetic data, and small 1B generators match or beat 27B generators for pretraining-data synthesis.
-
Stress-Testing the Reasoning Competence of LLMs With Proofs Under Minimal Formalism
ProofGrid is a new benchmark for LLM reasoning that uses machine-checkable proofs in minimal formal notation, revealing progress on basic tasks but major gaps in complex combinatorial and synthesis reasoning.
-
Improving LLM Unlearning Robustness via Random Perturbations
LLM unlearning is reframed as inadvertently installing backdoor triggers on forget-tokens; Random Noise Augmentation is introduced as a defense that improves robustness with theoretical guarantees.
-
DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model
DeepSeek-V2 delivers top-tier open-source LLM performance using only 21B active parameters by compressing the KV cache 93.3% and cutting training costs 42.5% via MLA and DeepSeekMoE.
-
RouterBench: A Benchmark for Multi-LLM Routing System
RouterBench supplies a standardized benchmark, 405k+ inference dataset, theoretical framework, and comparative analysis for multi-LLM routing systems.
-
Self-Rewarding Language Models
Iterative self-rewarding via LLM-as-Judge in DPO training on Llama 2 70B improves instruction following and self-evaluation, outperforming GPT-4 on AlpacaEval 2.0.
-
GAIA: a benchmark for General AI Assistants
GAIA benchmark shows humans at 92% accuracy on simple real-world questions far outperform current AI systems at 15%, proposing this gap as a key milestone for general AI.
-
Measuring Faithfulness in Chain-of-Thought Reasoning
Chain-of-Thought reasoning in LLMs is often unfaithful, with models relying on it variably by task and less so as models scale larger.
-
OPT: Open Pre-trained Transformer Language Models
OPT releases open decoder-only transformers up to 175B parameters that match GPT-3 performance at one-seventh the carbon cost, along with code and training logs.
-
SLORR: Simple and Efficient In-Training Low-Rank Regularization
A stateless, SVD-free regularizer approximates polar factors to induce low-rank weight structure during training, enabling better post-training compression of vision models and LLMs at under 8% overhead.
-
Understanding Layer Patching in Model Size Interpolation
Optimal layer-patching order for boomerang distillation is a shortest path on a KL-weighted Boolean lattice; greedy KLPatch and simple sequential orders often yield near-optimal interpolations.
-
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.
-
FlexMoE: One-for-All Nested Intra-Expert Pruning for MoE Language Models
FlexMoE produces nested pruned subnetworks for MoE LLMs across budgets via channel importance ranking and discrete action learning, plus one mid-budget recovery fine-tune, retaining 99.8% performance at 50% expert parameter pruning.
-
Aurora: A Leverage-Aware Spectral Optimizer
Aurora is a leverage-aware spectral optimizer that enforces uniform row norms in matrix updates while preserving Muon's polar geometry, outperforming Muon and achieving SOTA among spectral methods on modded-nanoGPT.
-
ConSA: Controllable Sparsity in Hybrid Attention via Learnable Allocation
ConSA learns FA/SWA allocation via L0 masks and augmented Lagrangian constraints, outperforming rule-based baselines on 0.6B and 1.7B models with consistent layer patterns.
-
M\"OVE: A Holistic LLM Benchmark for the German Public Sector
MÖVE presents a new German-language benchmark evaluating 39 LLMs on performance and governance criteria using ten public-administration datasets.
-
Redesign Mixture-of-Experts Routers with Manifold Power Iteration
Manifold Power Iteration aligns MoE router rows with principal singular directions of experts via a power-then-retract process, with theory showing convergence and experiments on 1B-11B models showing gains.
-
Compress-Distill: Reasoning Trace Compression for Efficient Knowledge Distillation
Post-hoc model-based compression of reasoning traces cuts training tokens to 12-30% and speeds training 2-7.6x while retaining up to 96% of raw-trace accuracy, though raw traces remain superior at every scale.
-
Dynamic Short Convolutions Improve Transformers
Dynamic short convolutions applied to key/query/value projections and linear layers in Transformers yield consistent performance gains and 1.33-1.60x compute advantages over standard models on language modeling from 150M to 2B parameters.
-
Scaling Agentic Capabilities via Grounded Interaction Synthesis
GAIS synthesizes diverse, high-fidelity agentic tasks from real-world MCP servers and adversarial planning, outperforming LLM-only baselines on BFCL, τ²-Bench, and ACEBench with greater data efficiency.
-
ScaleSweep: Accurate NVFP4 Post-Training Quantization of LLMs via Block Scale Initialization
ScaleSweep sweeps block scale candidates for NVFP4 quantization to minimize MSE/WMSE, derives bounds to shrink the search, and reports improved performance on Llama and Qwen models over AbsMax initialization.
-
Multilingual Knowledge Transfer under Data Constraints via Lexical Interventions
LINK improves cross-lingual knowledge transfer via lexical substitutions in English pretraining data, yielding notable downstream gains and up to 2x training speedup across eight languages and five model sizes.
-
How Human-Like Are Large Language Models? A Register-Aware Linguistic Evaluation Framework
A new evaluation framework using MMD on Biber features shows LLMs deviate from human linguistic distributions across registers, with closest models varying by register rather than size.
-
PALS: Power-Aware LLM Serving for Mixture-of-Experts Models
PALS adds dynamic GPU power capping to LLM serving frameworks like vLLM, jointly tuning it with batch size via offline models and feedback control to improve energy efficiency up to 26.3% and cut QoS violations 4-7x on dense and MoE models.
-
Flash PD-SSM: Memory-Optimized Structured Sparse State-Space Models
Flash PD-SSM achieves FSA-level expressivity by discretely selecting one matrix from a trainable set of structured sparse transition matrices at each time step while preserving the runtime and memory efficiency of standard structured SSMs.
-
Forecasting Downstream Performance of LLMs With Proxy Metrics
Proxy metrics from next-token distributions over expert solutions outperform loss and compute baselines for ranking LLMs, selecting pretraining data, and extrapolating performance across compute scales.
-
Scaling Laws for Mixture Pretraining Under Data Constraints
Empirical study shows mixture pretraining tolerates higher target data repetition than single-source training, with a new repetition-aware scaling law enabling principled mixture selection based on data size, compute, and model scale.
-
MuonQ: Enhancing Low-Bit Muon Quantization via Directional Fidelity Optimization
A 4-bit quantization scheme for the Muon optimizer — combining pre-quantization normalization, singular-structure-aligned decomposition, and µ-law companding — recovers most full-precision training quality with up to 7.3x optimizer-state memory savings.
-
ADMM-Q: An Improved Hessian-based Weight Quantizer for Post-Training Quantization of Large Language Models
ADMM-Q is a new post-training quantization method using ADMM operator splitting that reduces WikiText-2 perplexity compared to GPTQ on Qwen3-8B across W3A16, W4A8, and W2A4KV4 settings.
-
Training Transformers for KV Cache Compressibility
Training transformers with KV sparsification during continued pretraining produces representations that admit better post-hoc KV cache compression, improving quality under memory budgets for long-context tasks.
-
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.
-
Minimizing Collateral Damage in Activation Steering
Activation steering is cast as constrained optimization that minimizes collateral damage by weighting perturbations according to the empirical second-moment matrix of activations instead of assuming isotropy.
-
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.
-
SAMoRA: Semantic-Aware Mixture of LoRA Experts for Task-Adaptive Learning
SAMoRA is a parameter-efficient fine-tuning framework that uses semantic-aware routing and task-adaptive scaling within a Mixture of LoRA Experts to improve multi-task performance and generalization over prior methods.
-
MOONSHOT : A Framework for Multi-Objective Pruning of Vision and Large Language Models
MOONSHOT is a wrapper framework that combines reconstruction loss and Taylor approximation objectives to improve one-shot pruning results on LLMs and vision models.
-
ReSpinQuant: Efficient Layer-Wise LLM Quantization via Subspace Residual Rotation Approximation
Layer-wise LLM rotations can be fused offline and residual basis mismatch approximated by a low-rank subspace, matching expensive layer-wise PTQ accuracy with near-global overhead.
-
EdgeFlow: Fast Cold Starts for LLMs on Mobile Devices
EdgeFlow reduces mobile LLM cold-start latency up to 4.07x versus llama.cpp, MNN, and llm.npu by NPU-aware adaptive quantization, SIMD-friendly packing, and synergistic granular CPU-NPU pipelining at comparable accuracy.
-
Initialisation Determines the Basin: Efficient Codebook Optimisation for Extreme LLM Quantization
Output-aware EM initialization for codebooks in additive quantization avoids poor optimization basins and yields better 2-bit compressed LLMs across Llama and Qwen models.
-
TalkLoRA: Communication-Aware Mixture of Low-Rank Adaptation for Large Language Models
TalkLoRA equips MoE-LoRA experts with a communication module that smooths routing dynamics and improves performance on language tasks under similar parameter budgets.
-
LaCy: What Small Language Models Can and Should Learn is Not Just a Question of Loss
A small language model's factual accuracy in cascades improves when the decision to delegate a token is made from both token loss and spaCy factuality labels, not loss alone.
-
mHC: Manifold-Constrained Hyper-Connections
mHC projects hyper-connection residual spaces onto a manifold to restore identity mapping, enabling stable large-scale training with performance gains over standard HC.