A blank-image ablation test reveals that high probe accuracy on VLM spatial reasoning frequently reflects priors or inverted signs rather than image grounding, with horizontal grounded, vertical prior, and depth inverted.
super hub Mixed citations
Gemma 3 Technical Report
Mixed citation behavior. Most common role is background (70%).
572
Pith papers citing it
251
external citations · Pith
Background
70% of classified citations
abstract
We introduce Gemma 3, a multimodal addition to the Gemma family of lightweight open models, ranging in scale from 1 to 27 billion parameters. This version introduces vision understanding abilities, a wider coverage of languages and longer context - at least 128K tokens. We also change the architecture of the model to reduce the KV-cache memory that tends to explode with long context. This is achieved by increasing the ratio of local to global attention layers, and keeping the span on local attention short. The Gemma 3 models are trained with distillation and achieve superior performance to Gemma 2 for both pre-trained and instruction finetuned versions. In particular, our novel post-training recipe significantly improves the math, chat, instruction-following and multilingual abilities, making Gemma3-4B-IT competitive with Gemma2-27B-IT and Gemma3-27B-IT comparable to Gemini-1.5-Pro across benchmarks. We release all our models to the community.
hub tools
citation-role summary
background 50
baseline 10
other 4
dataset 3
method 2
citation-polarity summary
claims ledger
- abstract We introduce Gemma 3, a multimodal addition to the Gemma family of lightweight open models, ranging in scale from 1 to 27 billion parameters. This version introduces vision understanding abilities, a wider coverage of languages and longer context - at least 128K tokens. We also change the architecture of the model to reduce the KV-cache memory that tends to explode with long context. This is achieved by increasing the ratio of local to global attention layers, and keeping the span on local attention short. The Gemma 3 models are trained with distillation and achieve superior performance to Gem
authors
co-cited works
representative citing papers
Noisy expert imitation learning requires exponential samples for offline methods but polynomial for a variant of on-policy distillation under a noise condition.
MedCUA-Bench provides 18 clinical scenarios in 10 domains as a testbed for computer-use agents on medical UIs, with evaluations of 23 agents showing low success rates especially on real systems like OpenEMR.
Looped linear transformers with LN provably converge via GD to implement the power method on principal component prediction.
MM-JudgeBench shows substantial cross-lingual performance variance in 22 LVLM judges, with model size and architecture as poor predictors of multilingual robustness.
SAHM is the first Arabic financial benchmark with seven tasks including AAOIFI standards QA, fatwa reasoning, accounting exams, sentiment analysis, summarization, and event-cause reasoning, showing that Arabic fluency does not imply strong financial reasoning in 20 tested LLMs.
ArgBench unifies 33 existing datasets into a standardized benchmark for testing LLMs across 46 argumentation tasks and analyzes the impact of prompting techniques and model factors on performance.
VAREX benchmark shows structured output compliance limits models under 4B parameters more than extraction ability, with layout-preserving text giving the largest accuracy gains over images.
CELM is the first EEG-to-language foundation model that generates clinical reports from variable-length EEG recordings using a new dataset of 9,922 reports paired with 11,000 hours of data from 9,048 patients.
Molmo2 delivers state-of-the-art open-weight video VLMs with new grounding datasets and training methods that outperform prior open models and match or exceed some proprietary ones on pointing and tracking tasks.
Six state-of-the-art LLMs systematically prefer Standard American English over AAE continuations, and a training-free activation steering method reduces this bias 5-20x more than prompting while preserving fluency.
EAGOR reformulates embodied 360-degree directional reasoning as recursive Bayesian estimation on a spherical manifold using spherical harmonics, achieving training-free, rotation-equivariant target tracking.
MTLA is a training-free, post-hoc confidence score for multimodal LLM localization that restricts attention aggregation to the model's own predicted region and tokens, substantially improving hallucination detection and re-ranking across image, video, and audio.
NEvo performs evolutionary search guided by a dynamic voxel-level encoding model to synthesize videos that maximize predicted activity in target brain ROIs, recovering known selectivities and revealing temporal dynamics differences.
EgoSafetyBench shows VLMs reliably spot hazard-containing videos but miss specific contextual hazards and are degraded by misleading in-scene text.
Self-generated QA supervision for language models is fragile due to non-uniform question selection and instruction compliance during answering, with mitigations that reduce compliance from 88% to 13%.
RCT dataset with sequence-preserving splits demonstrates that tactile-to-text models achieve only 25.1% Recall@1 on held-out materials, exposing generalization as the core challenge.
VSE perturbs images only to probe visual ambiguity in VLMs, clusters outputs into semantic prototypes, and computes mass-weighted dispersion, outperforming prior entropy methods on five VQA benchmarks across five models.
LaViD distills LLM conceptual knowledge to vision models via LLM-generated MCQ soft labels, outperforming vision-language distillation baselines on fine-grained benchmarks while improving robustness on spurious correlation datasets.
DiCoBench is a new high-resolution multi-image benchmark exposing large gaps between top MLLMs and human performance (98.3%) on differential and commonality visual cue perception.
Cliff tokens are single tokens triggering LLM math reasoning failures, identified via adaptive z-test threshold on token potential; a taxonomy and Cliff-DPO optimization yield up to +6.6 accuracy gains.
PACE is a clipped per-coordinate controller added to AdamW that improves the limiting error of the returned iterate average in both quadratic analysis and LM experiments.
Forensic Knowledge Graphs integrate forensic traces, causal dependencies, and scene links via a new authentication network and Iterative Context Refinement to outperform standard detectors and VLMs on detection, localization, and justification.
FlexServe introduces recallable secure memory and NPU to enable cooperative secure LLM inference on mobile devices, reporting 10.05X TTFT speedup over a basic TrustZone strawman.
citing papers explorer
-
Decodable Is Not Grounded: A Vision-Ablation Arbiter for VLM Spatial Reasoning
A blank-image ablation test reveals that high probe accuracy on VLM spatial reasoning frequently reflects priors or inverted signs rather than image grounding, with horizontal grounded, vertical prior, and depth inverted.
-
Behavior Cloning is Not All You Need: The Optimality of On-Policy Distillation for Noisy Expert Feedback
Noisy expert imitation learning requires exponential samples for offline methods but polynomial for a variant of on-policy distillation under a noise condition.
-
MedCUA-Bench: A Screenshot-Only Benchmark for Clinical Computer-Use Agents
MedCUA-Bench provides 18 clinical scenarios in 10 domains as a testbed for computer-use agents on medical UIs, with evaluations of 23 agents showing low success rates especially on real systems like OpenEMR.
-
Looped Transformers with Layer Normalization Provably Learn the Power Method
Looped linear transformers with LN provably converge via GD to implement the power method on principal component prediction.
-
Lost in Translation: Do LVLM Judges Generalize Across Languages?
MM-JudgeBench shows substantial cross-lingual performance variance in 22 LVLM judges, with model size and architecture as poor predictors of multilingual robustness.
-
SAHM: A Benchmark for Arabic Financial and Shari'ah-Compliant Reasoning
SAHM is the first Arabic financial benchmark with seven tasks including AAOIFI standards QA, fatwa reasoning, accounting exams, sentiment analysis, summarization, and event-cause reasoning, showing that Arabic fluency does not imply strong financial reasoning in 20 tested LLMs.
-
ArgBench: Benchmarking LLMs on Computational Argumentation Tasks
ArgBench unifies 33 existing datasets into a standardized benchmark for testing LLMs across 46 argumentation tasks and analyzes the impact of prompting techniques and model factors on performance.
-
VAREX: A Benchmark for Multi-Modal Structured Extraction from Documents
VAREX benchmark shows structured output compliance limits models under 4B parameters more than extraction ability, with layout-preserving text giving the largest accuracy gains over images.
-
Neural Signals Generate Clinical Notes in the Wild
CELM is the first EEG-to-language foundation model that generates clinical reports from variable-length EEG recordings using a new dataset of 9,922 reports paired with 11,000 hours of data from 9,048 patients.
-
Molmo2: Open Weights and Data for Vision-Language Models with Video Understanding and Grounding
Molmo2 delivers state-of-the-art open-weight video VLMs with new grounding datasets and training methods that outperform prior open models and match or exceed some proprietary ones on pointing and tracking tasks.
-
LLMs Silently Correct African American English: Auditing and Mitigating Dialect Bias via Activation Steering
Six state-of-the-art LLMs systematically prefer Standard American English over AAE continuations, and a training-free activation steering method reduces this bias 5-20x more than prompting while preserving fluency.
-
EAGOR: Embodied Reasoning in Omni-direction
EAGOR reformulates embodied 360-degree directional reasoning as recursive Bayesian estimation on a spherical manifold using spherical harmonics, achieving training-free, rotation-equivariant target tracking.
-
Propose and Attend: Training-free MLLM Grounding Confidence via Multi-Token Localized Attention
MTLA is a training-free, post-hoc confidence score for multimodal LLM localization that restricts attention aggregation to the model's own predicted region and tokens, substantially improving hallucination detection and re-ranking across image, video, and audio.
-
NEvo: Neural-Guided Evolutionary Video Synthesis for Dynamic Visual Selectivity
NEvo performs evolutionary search guided by a dynamic voxel-level encoding model to synthesize videos that maximize predicted activity in target brain ROIs, recovering known selectivities and revealing temporal dynamics differences.
-
EgoSafetyBench: A Diagnostic Egocentric Video Benchmark for Evaluating Embodied VLMs as Runtime Safety Guards
EgoSafetyBench shows VLMs reliably spot hazard-containing videos but miss specific contextual hazards and are degraded by misleading in-scene text.
-
Self-Study Reconsidered: The Hidden Fragility of Learning from Self-Generated QA
Self-generated QA supervision for language models is fragile due to non-uniform question selection and instruction compliance during answering, with mitigations that reduce compliance from 88% to 13%.
-
RCT: A Robot-Collected Touch-Vision-Language Dataset for Tactile Generalization
RCT dataset with sequence-preserving splits demonstrates that tactile-to-text models achieve only 25.1% Recall@1 on held-out materials, exposing generalization as the core challenge.
-
Visual Semantic Entropy: Do Vision Language Models Recognize Visual Ambiguity?
VSE perturbs images only to probe visual ambiguity in VLMs, clusters outputs into semantic prototypes, and computes mass-weighted dispersion, outperforming prior entropy methods on five VQA benchmarks across five models.
-
Large Language Model Teaches Visual Students: Cross-Modality Transfer of Fine-Grained Conceptual Knowledge
LaViD distills LLM conceptual knowledge to vision models via LLM-generated MCQ soft labels, outperforming vision-language distillation baselines on fine-grained benchmarks while improving robustness on spurious correlation datasets.
-
DiCoBench: Benchmarking Multi-Image Fine-Grained Perception via Differential and Commonality Visual Cues
DiCoBench is a new high-resolution multi-image benchmark exposing large gaps between top MLLMs and human performance (98.3%) on differential and commonality visual cue perception.
-
Cliff Tokens: Identifying Single-Token Failure Triggers in LLM Mathematical Reasoning
Cliff tokens are single tokens triggering LLM math reasoning failures, identified via adaptive z-test threshold on token potential; a taxonomy and Cliff-DPO optimization yield up to +6.6 accuracy gains.
-
Training for the Model You Return: Improving Optimization for Iterate-Averaged Language Models
PACE is a clipped per-coordinate controller added to AdamW that improves the limiting error of the returned iterate average in both quadratic analysis and LM experiments.
-
Trustworthy Image Authentication using Forensic Knowledge Graphs
Forensic Knowledge Graphs integrate forensic traces, causal dependencies, and scene links via a new authentication network and Iterative Context Refinement to outperform standard detectors and VLMs on detection, localization, and justification.
-
FlexServe: A Fast and Secure LLM Serving System for Mobile Devices with Flexible Resource Isolation
FlexServe introduces recallable secure memory and NPU to enable cooperative secure LLM inference on mobile devices, reporting 10.05X TTFT speedup over a basic TrustZone strawman.
-
GitReq: A Gold Standard Dataset for Software Quality Requirements
GitReq is a released dataset of 6,302 expert-labeled GitHub requirements across eight ISO 25010 categories with LLM classification baselines.
-
FAPO: Fully Automated Prompt Optimization of Multi-Step LLM Pipelines
FAPO automates LLM pipeline optimization via iterative diagnosis and prompt-or-structure edits, beating GEPA baseline by +14.1 pp mean across 18 comparisons and +33.8 pp when structural changes occur.
-
Algebraic Dead Directions in LayerNorm Transformers: A Forward-Pass-Only Diagnostic at LLM Scale
The normalized inverse-scale direction of LayerNorm's affine parameters is an exact algebraic kernel of the post-final-norm centred activation covariance for any input distribution in LayerNorm transformers.
-
Image Prompt Reconstruction Attacks on Distributed MLLM Inference Frameworks
First study of image prompt reconstruction attacks on distributed MLLM inference, proposing MPAA for pixel-level and IEDA for semantic reconstruction with 100% embedding extraction accuracy on four model families.
-
TIGER: Inverting Transformer Gradients via Embedding-Subspace Distance Optimization
TIGER turns the low-rank attention gradient subspace into a differentiable objective for continuous embedding optimization, improving reconstruction quality and robustness over prior discrete token tests especially under noise or DP.
-
Steering Where to Listen: Instruction-Based Activation Steering Redirects Temporal Attention in Large Audio-Language Models
Instruction-based vector steering redirects temporal attention in LALMs to acoustically relevant regions, recovering queried sound event locations with 60.87-68.72% overlap accuracy without training.
-
A History-Aware Visually Grounded Critic for Computer Use Agents
HiViG is a test-time critic that combines macro-action history summarization with visual grounding of execution coordinates to reduce short-sighted and visually erroneous actions in long-horizon GUI agents.
-
Reason Twice: Segmentation via Candidate Discovery and Comparative Reasoning
Rea2Seg turns image segmentation into candidate mask discovery from MLLM attention followed by MLLM-based comparative scoring and selection, plus a new multi-dimensional reasoning benchmark ReasonSeg-SGDR.
-
When Correct Decisions Hide Internal Stress: Decision-State Probing in Multimodal Language Models
S³E framework finds excess decision-state displacement under semantic stress in multimodal models despite consistent correct forced-choice behavior.
-
When No Answer Is Correct: Diagnosing Absent Answer Detection for MLLMs in Video Understanding
MLLMs fail to detect absent correct answers in video QA tasks across three evaluation settings, defaulting to distractors even with chain-of-thought prompting.
-
Sci-Rho: A Multilingual Visually-Grounded Symbolic Benchmark for STEM Problems
Sci-Rho is a dynamic multilingual visually-grounded symbolic benchmark for STEM problems that reveals robustness gaps in current VLMs between average and worst-case performance.
-
UrduMMLU: A Massive Multitask Benchmark for Urdu Language Understanding
UrduMMLU is a new native-source MCQ benchmark for Urdu that reveals top LLMs reach only ~90% accuracy with large gaps on region-specific humanities content.
-
NAVI-Orbital: First In-Orbit Demonstration of a Zero-Shot Vision-Language Model for Autonomous Earth Observation
NAVI-Orbital performed the first claimed in-orbit demonstration of a zero-shot vision-language model for autonomous Earth observation using Gemma 3 on April 16, 2026.
-
Anchored, Not Graded: Vision-Language Models Fail at Slant-from-Texture Perception
VLMs across families and scales show anchoring to discrete slant angles in zero-shot and prompted settings rather than human-like graded texture-based slant perception.
-
MMBU: A Massive Multi-modal Biomedical Understanding Benchmark to Probe the Perception Capabilities of Vision-Language Models
Introduces MMBU benchmark for VLMs in biomedicine and demonstrates that established benchmarks mask perception deficiencies in evaluated models.
-
Why Muon Outperforms Adam: A Curvature Perspective
Muon outperforms Adam by reducing curvature penalty via lower Normalized Directional Sharpness, as shown via Taylor approximation on LLM training and proven on stylized quadratic problems with heterogeneous curvature.
-
Text-to-Image Models Need Less from Text Encoders Than You Think
A bag-of-position-tagged-words embedding guides text-to-image diffusion models as effectively as full contextual text embeddings from standard encoders.
-
SEA-NLI: Natural Language Inference as a Lens into Southeast Asian Cultural Understanding
SEA-NLI benchmark shows low performance across 17 LLMs on Southeast Asian cultural NLI, mainly due to missing cultural knowledge, with gains from SEA-adapted models and culture-aware prompting.
-
GeoDrive-Bench: Benchmarking Region-Specific Multimodal Reasoning in Autonomous Driving
GeoDrive-Bench is a new multimodal benchmark and distillation method for testing and improving VLMs on region-specific traffic-rule reasoning in autonomous driving across six countries.
-
Not What, But How: A Framework for Auditing LLM Responses across Positioning, Generalization, Anthromorphism, and Maxims
Presents FRANZ framework and SQUARE corpus for multi-dimensional audit of LLM response framing on subjective cultural queries, applied to three models to reveal differences and couplings.
-
Decentralized Instruction Tuning: Conflict-Aware Splitting and Weight Merging
MERIT enables decentralized instruction tuning via conflict-aware PCA splitting and parameter-space merging, raising average benchmark scores above joint training on multimodal and text mixtures.
-
Attention-guided Fine-tuning of Multimodal Large Language Models Improves Chain-of-Thought Reasoning
Attentive-CoT is an attention-guided fine-tuning objective that improves chain-of-thought performance in multimodal LLMs by delaying answer commitment and increasing sustained visual-token access during rationale generation.
-
OmniOPD: Logit-Free On-Policy Distillation via Speculative Verification
OmniOPD replaces token-level logit matching in on-policy distillation with Monte Carlo chunk-level semantic verification and a peak-entropy scheduler.
-
Subliminal Learning Is Steering Vector Distillation
Subliminal learning is steering vector distillation: a student fine-tuned on a steered teacher's outputs learns to imitate the steering vector.
-
How Far Do Auto-Interpretation Labels Generalize: A Controlled Study Across Languages, Scripts, and Rewordings
Auto-interpretation labels for SAE features generalize poorly across languages and scripts, missing the same semantic content up to 4x more often in Serbian than English and more in Cyrillic than Latin despite deterministic transliteration.
-
LLMs Need Encoders for Semantic IDs Too
PrefixMem encoder for Semantic IDs improves deepest-level accuracy by up to 46% relative and full-SID retrieval recall by up to 22% relative on Pinterest data across LLM families.