CiteVQA requires models to cite specific document regions with bounding boxes alongside answers and finds that even the strongest MLLMs frequently cite the wrong region, with top SAA scores of only 76.0 for closed models and 22.5 for open-source ones.
super hub
Gemma: Open Models Based on Gemini Research and Technology
105 Pith papers cite this work. Polarity classification is still indexing.
105
Pith papers citing it
abstract
This work introduces Gemma, a family of lightweight, state-of-the art open models built from the research and technology used to create Gemini models. Gemma models demonstrate strong performance across academic benchmarks for language understanding, reasoning, and safety. We release two sizes of models (2 billion and 7 billion parameters), and provide both pretrained and fine-tuned checkpoints. Gemma outperforms similarly sized open models on 11 out of 18 text-based tasks, and we present comprehensive evaluations of safety and responsibility aspects of the models, alongside a detailed description of model development. We believe the responsible release of LLMs is critical for improving the safety of frontier models, and for enabling the next wave of LLM innovations.
hub tools
citation-role summary
background 2
dataset 1
citation-polarity summary
claims ledger
- abstract This work introduces Gemma, a family of lightweight, state-of-the art open models built from the research and technology used to create Gemini models. Gemma models demonstrate strong performance across academic benchmarks for language understanding, reasoning, and safety. We release two sizes of models (2 billion and 7 billion parameters), and provide both pretrained and fine-tuned checkpoints. Gemma outperforms similarly sized open models on 11 out of 18 text-based tasks, and we present comprehensive evaluations of safety and responsibility aspects of the models, alongside a detailed descript
authors
co-cited works
representative citing papers
LLM popularity judgments align more closely with pretraining data exposure counts than with Wikipedia popularity, with stronger effects in pairwise comparisons and larger models.
A classifier trained only on transformer fine-tuning data detects an invariant memorization signature that transfers to Mamba, RWKV-4, and RecurrentGemma with AUCs of 0.963, 0.972, and 0.936.
MiCP is the first conformal prediction method for multi-turn LLM pipelines that allocates per-turn error budgets to enable adaptive stopping with an overall coverage guarantee, shown to reduce turns and cost on RAG and ReAct benchmarks.
Linear representations of high-level concepts in LLMs are formalized via counterfactuals in input and output spaces, unified under a causal inner product that enables consistent probing and steering.
A new speculative inference system speeds up diffusion VLAs to 19.1 ms average latency (3.04x faster) on LIBERO by replacing most full 58 ms inferences with 7.8 ms draft rounds while preserving task performance.
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.
A parallel multi-turn medical dialogue dataset spanning English and nine Indic languages is created from synthetic consultations to enable personalized AI healthcare interactions.
LiteLVLM prunes visual tokens for pixel grounding by reversing CLIP visual-text similarity to retain referent region tokens, outperforming prior methods by over 5% with 22% speedup and 2.3x memory reduction without any training.
Projecting LLM hidden states onto F2 algebra with 42 pairs yields 93% zero-shot accuracy on logical relations and identifies prompt-preventable late-layer collapse.
Chronicles-OCR is the first benchmark with 2,800 images across the complete evolutionary trajectory of Chinese characters, defining four tasks to evaluate VLLMs' cross-temporal visual perception.
Chirality emerges in SMILES translation models through an abrupt encoder-centered reorganization of representations after a long plateau, identified via checkpoint analysis and ablation.
EDEN adaptively sets branching factor proportional to next-token entropy, achieving better accuracy per expansion than fixed beam search while providing a proof that monotone entropy-based branching outperforms any fixed budget allocation.
K12-KGraph is a textbook-derived knowledge graph that powers a new benchmark revealing LLMs' poor curriculum cognition and a small training corpus that outperforms general instruction data on educational tasks.
PAS encodes locations via relative anchors and bins to deliver roughly 370-400m adversarial error in spatial RAG while retaining over half the baseline retrieval performance and keeping generation quality robust.
Black-box membership inference attacks on retrieval-based in-context learning for document QA succeed via query prefixes, with a novel weighted-averaging method outperforming priors even under paraphrasing.
Decoding-time use of process reward models for bias mitigation raises fairness scores by up to 0.40 on a bilingual benchmark while preserving fluency across four LLMs and extends to open-ended generation with low overhead.
RAGCharacter localizes poisoned character spans in RAG evidence via prompt-conditioned counterfactual masking and achieves the best accuracy-over-attribution trade-off across tested attacks and models.
Full development of 7B and 32B Olmo 3 models used 12.3 GWh datacenter energy and emitted 4,251 tCO2eq, with development overheads accounting for 82% of compute and reasoning models costing 17x more to post-train than instruction-tuned ones.
Tempus delivers 607 GOPS at 10.677 W using fixed 16 AIE cores on Versal AI Edge, with 211.2x better platform-aware utility than spatial SOTA ARIES and zero URAM/DSP utilization.
ARA jailbreaks safety-aligned LLMs like LLaMA-3 and Mistral by redirecting attention in safety-heavy heads with as few as 5 tokens, achieving 30-36% attack success while ablating the same heads barely affects refusals.
Incompressible Knowledge Probes enable log-linear estimation of LLM parameter counts from factual accuracy on obscure questions, showing continued scaling of knowledge capacity across open and closed models.
Skip-connected MLPs and residual-free MLPs of equal width represent generically disjoint function classes for common activations, with explicit impossibility proofs and a non-generic absorption condition for ReLU and GELU.
Fine-tuning shows higher proficiency than in-context learning on in-distribution generalization in formal languages, with equal out-of-distribution performance and diverging inductive biases at high proficiency.
citing papers explorer
-
CiteVQA: Benchmarking Evidence Attribution for Trustworthy Document Intelligence
CiteVQA requires models to cite specific document regions with bounding boxes alongside answers and finds that even the strongest MLLMs frequently cite the wrong region, with top SAA scores of only 76.0 for closed models and 22.5 for open-source ones.
-
Pretraining Exposure Explains Popularity Judgments in Large Language Models
LLM popularity judgments align more closely with pretraining data exposure counts than with Wikipedia popularity, with stronger effects in pairwise comparisons and larger models.
-
Learning the Signature of Memorization in Autoregressive Language Models
A classifier trained only on transformer fine-tuning data detects an invariant memorization signature that transfers to Mamba, RWKV-4, and RecurrentGemma with AUCs of 0.963, 0.972, and 0.936.
-
Adaptive Stopping for Multi-Turn LLM Reasoning
MiCP is the first conformal prediction method for multi-turn LLM pipelines that allocates per-turn error budgets to enable adaptive stopping with an overall coverage guarantee, shown to reduce turns and cost on RAG and ReAct benchmarks.
-
The Linear Representation Hypothesis and the Geometry of Large Language Models
Linear representations of high-level concepts in LLMs are formalized via counterfactuals in input and output spaces, unified under a causal inner product that enables consistent probing and steering.
-
Realtime-VLA FLASH: Speculative Inference Framework for Diffusion-based VLAs
A new speculative inference system speeds up diffusion VLAs to 19.1 ms average latency (3.04x faster) on LIBERO by replacing most full 58 ms inferences with 7.8 ms draft rounds while preserving task performance.
-
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.
-
IndicMedDialog: A Parallel Multi-Turn Medical Dialogue Dataset for Accessible Healthcare in Indic Languages
A parallel multi-turn medical dialogue dataset spanning English and nine Indic languages is created from synthetic consultations to enable personalized AI healthcare interactions.
-
CLIP Tricks You: Training-free Token Pruning for Efficient Pixel Grounding in Large VIsion-Language Models
LiteLVLM prunes visual tokens for pixel grounding by reversing CLIP visual-text similarity to retain referent region tokens, outperforming prior methods by over 5% with 22% speedup and 2.3x memory reduction without any training.
-
Controlling Logical Collapse in LLMs via Algebraic Ontology Projection over F2
Projecting LLM hidden states onto F2 algebra with 42 pairs yields 93% zero-shot accuracy on logical relations and identifies prompt-preventable late-layer collapse.
-
Chronicles-OCR: A Cross-Temporal Perception Benchmark for the Evolutionary Trajectory of Chinese Characters
Chronicles-OCR is the first benchmark with 2,800 images across the complete evolutionary trajectory of Chinese characters, defining four tasks to evaluate VLLMs' cross-temporal visual perception.
-
From Syntax to Semantics: Unveiling the Emergence of Chirality in SMILES Translation Models
Chirality emerges in SMILES translation models through an abrupt encoder-centered reorganization of representations after a long plateau, identified via checkpoint analysis and ablation.
-
Entropy-informed Decoding: Adaptive Information-Driven Branching
EDEN adaptively sets branching factor proportional to next-token entropy, achieving better accuracy per expansion than fixed beam search while providing a proof that monotone entropy-based branching outperforms any fixed budget allocation.
-
K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs
K12-KGraph is a textbook-derived knowledge graph that powers a new benchmark revealing LLMs' poor curriculum cognition and a small training corpus that outperforms general instruction data on educational tasks.
-
Privacy Without Losing Place: A Paradigm for Private Retrieval in Spatial RAGs
PAS encodes locations via relative anchors and bins to deliver roughly 370-400m adversarial error in spatial RAG while retaining over half the baseline retrieval performance and keeping generation quality robust.
-
Membership Inference Attacks for Retrieval Based In-Context Learning for Document Question Answering
Black-box membership inference attacks on retrieval-based in-context learning for document QA succeed via query prefixes, with a novel weighted-averaging method outperforming priors even under paraphrasing.
-
Decoding-Time Debiasing via Process Reward Models: From Controlled Fill-in to Open-Ended Generation
Decoding-time use of process reward models for bias mitigation raises fairness scores by up to 0.40 on a bilingual benchmark while preserving fluency across four LLMs and extends to open-ended generation with low overhead.
-
Needle-in-RAG: Prompt-Conditioned Character-Level Traceback of Poisoned Spans in Retrieved Evidence
RAGCharacter localizes poisoned character spans in RAG evidence via prompt-conditioned counterfactual masking and achieves the best accuracy-over-attribution trade-off across tested attacks and models.
-
The Hidden Cost of Thinking: Energy Use and Environmental Impact of LMs Beyond Pretraining
Full development of 7B and 32B Olmo 3 models used 12.3 GWh datacenter energy and emitted 4,251 tCO2eq, with development overheads accounting for 82% of compute and reasoning models costing 17x more to post-train than instruction-tuned ones.
-
Tempus: A Temporally Scalable Resource-Invariant GEMM Streaming Framework for Versal AI Edge
Tempus delivers 607 GOPS at 10.677 W using fixed 16 AIE cores on Versal AI Edge, with 211.2x better platform-aware utility than spatial SOTA ARIES and zero URAM/DSP utilization.
-
Attention Is Where You Attack
ARA jailbreaks safety-aligned LLMs like LLaMA-3 and Mistral by redirecting attention in safety-heavy heads with as few as 5 tokens, achieving 30-36% attack success while ablating the same heads barely affects refusals.
-
Incompressible Knowledge Probes: Estimating Black-Box LLM Parameter Counts via Factual Capacity
Incompressible Knowledge Probes enable log-linear estimation of LLM parameter counts from factual accuracy on obscure questions, showing continued scaling of knowledge capacity across open and closed models.
-
Can an MLP Absorb Its Own Skip Connection?
Skip-connected MLPs and residual-free MLPs of equal width represent generically disjoint function classes for common activations, with explicit impossibility proofs and a non-generic absorption condition for ReLU and GELU.
-
Fine-tuning vs. In-context Learning in Large Language Models: A Formal Language Learning Perspective
Fine-tuning shows higher proficiency than in-context learning on in-distribution generalization in formal languages, with equal out-of-distribution performance and diverging inductive biases at high proficiency.
-
ParetoSlider: Diffusion Models Post-Training for Continuous Reward Control
ParetoSlider conditions diffusion models on continuous preference weights to approximate the full Pareto front, providing dynamic control over multi-objective rewards at inference time.
-
RSRCC: A Remote Sensing Regional Change Comprehension Benchmark Constructed via Retrieval-Augmented Best-of-N Ranking
RSRCC is a new 126k-question benchmark for fine-grained remote sensing change question-answering, constructed via a hierarchical semi-supervised pipeline with retrieval-augmented Best-of-N ranking.
-
How Tokenization Limits Phonological Knowledge Representation in Language Models and How to Improve Them
Subword tokenization impairs phonological knowledge encoding in LMs, but an IPA-based fine-tuning method restores it with minimal impact on other capabilities.
-
Understanding and Improving Continuous Adversarial Training for LLMs via In-context Learning Theory
Continuous adversarial training in the embedding space produces a robust generalization bound for linear transformers that decreases with perturbation radius, tied to singular values of the embedding matrix, and motivates a new regularizer that improves real LLM jailbreak robustness-utility tradeoff
-
TaxPraBen: A Scalable Benchmark for Structured Evaluation of LLMs in Chinese Real-World Tax Practice
TaxPraBen is a new benchmark with 14 datasets and a structured evaluation method for measuring LLM performance on Chinese real-world tax tasks and scenarios.
-
AtlasOCR: Building the First Open-Source Darija OCR Model with Vision Language Models
AtlasOCR delivers the first open-source Darija OCR by fine-tuning Qwen2.5-VL 3B, achieving state-of-the-art results on custom and existing benchmarks for both Darija and Arabic.
-
BOSCH: Black-Box Binary Optimization for Short-Context Attention-Head Selection in LLMs
BOSCH decomposes attention-head selection for short-context hybridization into layer probing, adaptive ratio assignment, and grouped binary optimization, yielding better efficiency-performance tradeoffs than static or layer-wise baselines.
-
Rank, Don't Generate: Statement-level Ranking for Explainable Recommendation
The work reframes explainable recommendation as statement-level ranking, introduces the StaR benchmark from Amazon reviews, and finds popularity baselines outperforming SOTA models in item-level personalized ranking.
-
Inside the Scaffold: A Source-Code Taxonomy of Coding Agent Architectures
Analysis of 13 coding agent scaffolds at pinned commits yields a 12-dimension taxonomy showing five composable loop primitives, with 11 agents combining multiple primitives instead of using one fixed structure.
-
The limits of bio-molecular modeling with large language models : a cross-scale evaluation
LLMs perform adequately on bio-molecular classification tasks but remain weak on regression, with hybrid architectures outperforming others on long sequences and fine-tuning hurting generalization.
-
Refusal in Language Models Is Mediated by a Single Direction
Refusal in language models is mediated by a single direction in residual stream activations that can be erased to disable safety or added to elicit refusal.
-
Jamba: A Hybrid Transformer-Mamba Language Model
Jamba presents a hybrid Transformer-Mamba MoE architecture for LLMs that delivers state-of-the-art benchmark performance and strong results up to 256K token contexts while fitting in one 80GB GPU with high throughput.
-
Learning to See What You Need: Gaze Attention for Multimodal Large Language Models
Gaze Attention groups visual embeddings into selectable regions and dynamically restricts attention to task-relevant ones, matching dense baselines with up to 90% fewer visual KV entries via added context tokens.
-
Large Spectrum Models (LSMs): Decoder-Only Transformer-Powered Spectrum Activity Forecasting via Tokenized RF Data
Decoder-only transformers trained on tokenized RF spectrum data from 22 TB of measurements achieve 3.25 dB RMSE in spectrum activity forecasting across 33 bands.
-
Remember to Forget: Gated Adaptive Positional Encoding
GAPE augments RoPE with query- and key-dependent gates to stabilize attention and improve long-context performance in language models.
-
Benchmarking Safety Risks of Knowledge-Intensive Reasoning under Malicious Knowledge Editing
EditRisk-Bench demonstrates that malicious knowledge editing reliably induces incorrect or unsafe reasoning in LLMs while largely preserving general capabilities.
-
ReST-KV: Robust KV Cache Eviction with Layer-wise Output Reconstruction and Spatial-Temporal Smoothing
ReST-KV formulates KV eviction as layer-wise output reconstruction optimization with spatial-temporal smoothing, outperforming baselines by 2.58% on LongBench and 15.2% on RULER while cutting decoding latency by 10.61x at 128k context.
-
Tool Calling is Linearly Readable and Steerable in Language Models
Tool identity is linearly readable and steerable in LLMs via mean activation differences, with 77-100% switch accuracy and error prediction from activation gaps.
-
ReasonEdit: Towards Interpretable Image Editing Evaluation via Reinforcement Learning
ReasonEdit uses a new CoT dataset and reinforcement learning to produce interpretable, human-aligned evaluations of text-guided image edits.
-
ModelLens: Finding the Best for Your Task from Myriads of Models
ModelLens learns a performance-aware latent space from 1.62M leaderboard records to rank unseen models on unseen datasets without forward passes on the target.
-
ChartZero: Synthetic Priors Enable Zero Shot Chart Data Extraction
ChartZero achieves zero-shot line chart data extraction by training only on synthetic mathematical functions, using a Global Orthogonal Instance loss to prevent curve fragmentation and a VLM-guided strategy for legend matching.
-
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.
-
RefusalGuard: Geometry-Preserving Fine-Tuning for Safety in LLMs
RefusalGuard constrains updates in hidden representation space to preserve safety-relevant geometric structure during fine-tuning, maintaining low attack success rates on safety benchmarks while preserving task performance.
-
Perturbation Probing: A Two-Pass-per-Prompt Diagnostic for FFN Behavioral Circuits in Aligned LLMs
Perturbation probing identifies tiny sets of FFN neurons that control refusal templates and language routing in LLMs, enabling precise ablations and directional interventions that alter behavior on benchmarks while preserving safety.
-
DUAL-BLADE: Dual-Path NVMe-Direct KV-Cache Offloading for Edge LLM Inference
DUAL-BLADE uses a dual-path KV-cache framework with NVMe-direct access to reduce prefill and decode latency by up to 33% and 42% while improving SSD utilization 2.2x under tight memory budgets.
-
The Override Gap: A Magnitude Account of Knowledge Conflict Failure in Hypernetwork-Based Instant LLM Adaptation
Knowledge conflicts in hypernetwork LLM adaptation stem from constant adapter margins losing to frequency-dependent pretrained margins; selective layer boosting and conflict-aware triggering raise deep-conflict accuracy to 71-72.5% on Gemma-2B and Mistral-7B.