Research Papers Compare
Head-to-head comparisons of methods and models, with numbers pulled from the source papers.
AI Agents · AdaPlanBench vs ToolMaze vs TASTE / τᶜ-Bench vs K-BrowseComp vs stale-observation masking
Four 2026 benchmarks stress-test agents: AdaPlanBench's hidden constraints (67.75% best), ToolMaze's silent tool failures (−37% PRR).
AI Agents · Eywa vs LabVLA vs Crafter
Each system owns one segment of the scientific loop: Eywa calls domain foundation models (0.6154 to 0.6558 utility), LabVLA runs lab protocols (71.1% ID), Crafter drafts editable figures (50.34 vs 11.13).
AI Agents · DRIFT vs Arbor
DRIFT finds the span where a deep-research trajectory breaks (54.91% macro-F1, up to 30 points over raw reads); Arbor blocks bad branches with hypothesis trees and a held-out gate (86.36% Any Medal on MLE-Bench Lite).
AI Agents · Pi-Bench vs SoCRATES vs TIDE
Three 2026 benchmarks measure three senses of 'proactive': Pi-Bench scores hidden-intent anticipation (best 67.0%), SoCRATES mediation (best closes a third of the gap), TIDE problem discovery (+20 over baselines).
Multimodal Models · MM-OCEAN vs Robust-U1
MM-OCEAN measures how often MLLMs answer correctly for the wrong visual reasons; Robust-U1 trains them to recover corrupted pixels before answering. One is a diagnostic, the other a repair.
Reinforcement Learning · Z-Reward vs CHERRL
Z-Reward predicts a full score distribution instead of a scalar to build a better rubric judge; CHERRL shows any rubric judge gets gamed inside a GRPO loop. Build one, guard with the other.
Efficient AI · LongLive-2.0 vs MinT
LongLive-2.0 runs a whole long-video model in NVFP4 4-bit, cutting a 64s generation from 112.9 to 36.3 GB; MinT swaps LoRA adapters over one resident base, making handoff 18.3x faster on a 4B dense LLM.
Agent Memory · MemDreamer vs TaskMem
MemDreamer fixes the read side with a training-free graph memory and agentic retrieval; TaskMem fixes the write side with an RL-trained memory policy adapted by a 2,048-parameter adapter.
Reinforcement Learning · DRPO vs Local Perturbation Theory
DRPO swaps DPPO's hard mask for a smooth Binary-TV weight, stable in 6/6 settings incl. FP8; Local Perturbation Theory shows multi-domain RL forgets via a second-order conflict subspace a short refresh repairs.
Text-to-Image · Z-Reward vs InterleaveThinker
Z-Reward replaces the scalar image reward with a score distribution (9B student at 88.6% accuracy, one token); InterleaveThinker wraps a frozen generator in a planner-critic loop (UEval 18.2 to 66.3).
World Models · Stream-R1 vs Stream-T1 vs WorldCraft
Stream-R1 reweights the distillation loss, Stream-T1 searches at inference, WorldCraft adds drag-and-drop object control: three budgets for the same streaming video stack, compared with numbers.
Vision Foundation Models · VLM3 vs ZipSplat vs Function2Scene vs DIRECT
VLM3 reads geometry from photos with a plain 4B VLM, ZipSplat rebuilds renderable scenes with 6-33x fewer Gaussians, Function2Scene turns a design brief into a layout.
LLM Reasoning · DeepSeek-R1 vs SU-01
DeepSeek-R1 proved RL alone can teach reasoning; SU-01 shows 338K short SFT traces plus 200 RL steps take a 30B open model to gold-level olympiad scores. What each recipe costs and when to pick which.
Mixture of Experts · MPI vs DAR vs dMoE
Three 2026 papers rebuild transformer routing: MPI lifts 11B MoE avg accuracy 40.92 to 42.76, DAR cuts SiT-XL/2 FID 9.67 to 7.56 using 8.75x fewer iterations, dMoE shrinks blocks from 69.5 to 14.6 experts.
Text-to-Image · Moebius vs Bootstrap Your Generator (ByG)
Moebius reaches FLUX-level inpainting at 0.22B via lambda blocks plus latent distillation; ByG trains image and video editors with zero paired data, winning 75.3% of video votes vs Ditto.
Reinforcement Learning · AntiSD vs SDPG vs SDAR
AntiSD rewards divergence from a privileged teacher (up to +11.5 points, 2-10x fewer steps than GRPO); SDPG distills an answer-hinted teacher (0.327 vs 0.242 AIME25); SDAR gates per-token trust (+10.2 WebShop).
Efficient AI · HRM-Text vs Lens vs Chinchilla
HRM-Text bets on architecture plus objective, Lens on data density, Chinchilla on budget reallocation. Same goal of cutting training cost, three levers, three failure modes.
Language Models · MTP vs NITP
MTP adds discrete heads that draft several tokens at once for throughput; NITP adds a dense target predicting the next token's shallow embedding for accuracy at ~2% extra training FLOPs.
AI Agents · Agents' Last Exam vs Pi-Bench
ALE tests whether agents finish explicit professional work (1,490 verifiable tasks, 2.6% pass on the hardest tier); Pi-Bench tests whether assistants catch what users never said (best proactivity 67.0%).
AI Agents · Role-Agent vs Skill1
Both train Qwen2.5 agents with GRPO-family RL on ALFWorld and WebShop and beat the same GiGPO baseline. Role-Agent makes the model its own reward and curriculum; Skill1 co-evolves a skill library.
Video Generation · OmniDirector vs SCAIL-2 vs Lip Forcing
OmniDirector, SCAIL-2, and Lip Forcing control camera, body, and lips in video diffusion, each escaping its axis's wall: paired data, pose skeletons, or bidirectional latency.
Multimodal Models · Flamingo vs LLaVA
Flamingo froze both backbones and learned image tasks from 32 in-context examples; LLaVA fine-tuned Vicuna on 158K GPT-4-written instructions. Same frozen-encoder trick, opposite task interface.
Vision Foundation Models · CLIP vs DINOv2
CLIP learns from 400M image-text pairs and classifies zero-shot via text prompts; DINOv2 uses no labels on 142M curated images and ships frozen features that beat OpenCLIP on most benchmarks.
Vision Foundation Models · SAM vs SAM 2
Meta's SAM 2 beats SAM on the same image suites (58.9 vs 58.1 one-click mIoU) while running 6x faster, and adds a streaming memory bank so a single prompt tracks an object through a whole video.
Open Models · Mistral 7B vs Llama 2
Mistral 7B beat Llama 2 13B on every benchmark its authors ran, at half the size, with grouped-query and sliding-window attention under Apache 2.0; Llama 2 adds a 70B tier and the first documented RLHF chat recipe.
Vision-Language-Action · PhysBrain 1.0 vs RATs vs RLDX-1
PhysBrain mines physics QA from human video (Franka grasping 47.1% to 63.3%), RATs banks play-generated code skills (LIBERO-PRO 23.2% to 43.8%), RLDX-1 adds motion, memory and tactile streams (catching 87.5% vs 29.2%).
Video Generation · Echo-Infinity vs MIGA
Echo-Infinity learns what to remember: 24-hour rollouts past 1.3M frames at 18.5 FPS on one H100, VBench-Long 82.01 at 240s; MIGA changes only the sampler and lifts VideoCrafter2 to 97.82 at 1000+ frames.
Video Generation · Stream-T1 vs VLM teacher
Stream-T1 searches a frozen streaming model's chunks at inference, lifting 5s VideoAlign motion quality from 0.350 to 0.629; the VLM-teacher loop optimizes a per-instance LoRA and lifts RULER-Bench from 46.4 to 68.2.
Diffusion Models · MV-Split (MMS) vs DAR
Two 2026 papers agree the inherited residual stream was never designed for diffusion: MV-Split cures deep-layer collapse (1000 layers, FID 2.77), DAR cures magnitude inflation (FID 9.67 to 7.56, 8.75x fewer iterations).
World Models · Echo-Memory vs Mirage
Echo-Memory isolates the memory module: block-wise recurrence hits 69.0 revisit consistency vs 12.25 with no memory. Mirage moves memory to latent space: 10.57x faster, 55x less GPU memory.
Long Context · Long Context vs KV Cache Sparsification vs Latent Context Compression
Gemini 1.5 scales the native window past 10M tokens; FlashMemory-DeepSeek-V4 keeps 13.5% of the KV cache with a lookahead indexer; LCLM compresses context into soft tokens at up to 1:16.
Agent Memory · MemPrivacy vs TaskMem vs persistent AI agents survey
MemPrivacy filters what may leave your device for cloud memory, holding utility loss within 1.6%. TaskMem learns what a video agent should write down at all, lifting VideoMME from 61.6% to 67.9%.
Small Language Models · BERT vs DistilBERT vs MobileBERT
DistilBERT halved BERT's depth and kept about 97% of its GLUE score; MobileBERT thinned it to a quarter of the size and ran 5.5x faster for a 0.6-point GLUE drop. Three papers, three compression strategies.
Efficient AI · Test-Time Training (TTT layer) vs Sleep (offline self-modification)
SD3.5-T5 uses a TTT layer's per-token gradient step to convert Softmax attention for about 1 hour, hitting 1.32x-1.47x speedup; Sleep shows test-time training fails at 10% ARC and moves learning offline to 80%.
Language Models · Chinchilla vs GPT-3
GPT-3 bet 175B parameters on roughly 300B tokens. Chinchilla re-ran the scaling math, showed data and parameters deserve equal scaling, and beat GPT-3 with a 70B model fed about 1.4T tokens under the same compute.
Vision Foundation Models · ResNet vs ViT vs CLIP
ResNet made depth trainable with skip connections; ViT beat CNNs once pre-trained on hundreds of millions of images; CLIP matched ResNet-50 zero-shot. Data scale, not architecture, decides the winner.
Vision-Language-Action · ACE-Ego-0 vs Humanoid-GPT vs DragMesh-2
ACE-Ego-0 imports human egocentric video into VLA pretraining, Humanoid-GPT makes 2 billion motion frames and distills 384 RL experts into one model, and DragMesh-2 strips actuation shortcuts.
Video Generation · SANA-Streaming vs LoomVideo vs CoVEBench
SANA-Streaming edits 1280x704 at 24 FPS on one RTX 5090; LoomVideo unifies generation and editing in 5B (VBench 63.15, 5.41x faster edits); CoVEBench shows editors still fail compositional instructions.
AI Agents · Video2GUI vs WeaveBench
Video2GUI mines 12M GUI trajectories from 500M web videos (+5-20% on Qwen2.5-VL, Mimo-VL); WeaveBench shows the best frontier pairing still fails 59% of 114 audited hybrid tasks.
LLM Reasoning · Chain-of-Thought Prompting vs ReAct vs Toolformer
CoT writes reasoning into the prompt, ReAct grounds that reasoning in tool observations, and Toolformer trains the model to call APIs itself. Same problem, three bets on where reliability comes from.
AI Agents · AutoResearchClaw vs EurekAgent vs ARIS vs ResearchClawBench
Four 2026 systems attack the same problem from different layers: workflow (0.648 vs 0.419), environment (85.71% any-medal on MLE-Bench Lite), adversarial assurance, and measurement.
Multimodal Models · SwanSphere vs Audio Interaction Model vs THUD
Three 2026 papers define real-time audio AI: SwanSphere streams spatial audio with a 0.21s first chunk, the Audio Interaction Model learns when to speak, and THUD catches models faking it.
Retrieval-Augmented Generation · GrepSeek vs FORT-Searcher vs SAAS
GrepSeek changes what the agent searches with (grep, no vector index), FORT-Searcher changes the training data (shortcut-resistant tasks), and SAAS changes when it stops (half the retrievals, equal accuracy).
Multimodal Models · X-Stream vs CARVE / V-RAGBench vs Watch-Remember-Reason survey
X-Stream tests watching concurrent live streams (best model 49.60% vs 91.84% human); V-RAGBench/CARVE tests retrieval over 11-99h egocentric video (Recall@5 0.603 vs 0.510).
AI Agents · Orchestra-o1 vs SearchSwarm vs StreamMA
Three June-2026 papers each fix one axis of a multi-agent LLM system: Orchestra-o1 decides who does what, SearchSwarm trains one model to delegate, StreamMA decides when agents exchange context.
Reinforcement Learning · CPPO vs TrOPD vs TRB
CPPO tightens early-token clipping in RL (+5.56 AIME over DPPO), TrOPD gates per-token teacher trust in distillation (+3.06 to +3.52), and TRB blends teacher rollouts inside a KL trust region during warmup.
Biomolecular Modeling · AlphaFold 3 vs ESM3 vs ProGen2 vs DynamicMPNN vs Feynman-Kac steering
Protein design is five different tasks wearing one name: AlphaFold 3 predicts structures, ESM3 and ProGen2 generate sequences, DynamicMPNN does inverse folding, and Feynman-Kac steering drives a diffusion designer.
Multimodal Models · ARM vs SenseNova-U1 vs HYDRA-X
ARM (7B, discrete tokens) reaches GenEval 0.86 after RL; SenseNova-U1 hits 0.91 GenEval and 80.55 MMMU; HYDRA-X extends one-tokenizer unification to video (GEdit 7.17).
World Models · Interactive world models vs World models for agents
Kairos runs a 4B world model on one RTX5090 (PAI-Bench 80.84); WALL-WM grounds robot actions in events (75.86 vs pi0.5's 55.64); Gamma-World runs four players at 24 FPS.
Multimodal Models · Spatial reasoning vs MLLMs vs visual agents
Spatial reasoning for MLLMs on three axes: SpatialClaw's Python workspace (59.9% on 20 benchmarks), ReRe's synthesized second view (+8.5 VSI-Bench on 2B), and frozen-probe evidence video models carry geometry VLMs lack.
Video Generation · Causal Forcing++ vs AnyFlow
Both distill Wan2.1 video diffusion for cheap sampling. Causal Forcing++ pins 1-2 steps per frame for 14.1 FPS real-time streaming; AnyFlow keeps steps free so quality scales from 4 to 32 NFE.
Speech Recognition · Whisper vs Mega-ASR
Whisper bets 680,000 hours of messy web audio on one robust generalist; Mega-ASR synthesizes 2.4M degraded clips to cut WER to 45.69% vs 54.01% on VOiCES. The numbers are not on the same exam.
Text-to-Image · Qwen-Image 2.0 vs Qwen-Image-Flash
Qwen-Image 2.0 is Alibaba's full-capability image generator and editor: 1K-token instructions, native 2K, a 16x VAE. Qwen-Image-Flash distills it to 4 steps, and the recipe lets the student match its teacher.
Brain Decoding · MindEye vs Brain Diffuser vs MinD-Vis
Same NSD benchmark, three systems: MindEye dominates retrieval at 93.6% against 300 candidates, Brain Diffuser wins SSIM, and MinD-Vis reports a different dataset entirely.
Text-to-Image · DALL·E 2 vs Stable Diffusion vs Imagen
DALL·E 2 diffuses in CLIP-embedding space with a learned prior, Stable Diffusion in a cheap latent, Imagen behind a frozen T5, and SD3 swaps the U-Net for a rectified-flow transformer.
Long Context · Kimi K3 vs DeepSeek V4 vs Qwen3.8-Next vs Nemotron 3 Ultra
Four 2026 frontier open models target 1M context four ways: Kimi K3 runs 3:1 delta attention, DeepSeek V4 compresses sparse KV, Qwen3.8-Next runs 3:1 GDN, Nemotron 3 Ultra leans on Mamba-2.
Multimodal Models · Keye-VL 2.0 vs InternVideo3
Keye-VL 2.0 buys long video with sparse-attention infrastructure: 30B-A3B, 256K context, 74.1 on LongVideoBench. InternVideo3 bets on a dense 8B plus inference-time tool use for +2.7 on Video-MME.
Reinforcement Learning · Agentic RL vs APPO vs LongTraceRL vs SDAR vs SCOPE
GRPO gives a multi-turn agent one scalar reward per episode. APPO, LongTraceRL, SDAR and SCOPE fix credit assignment at four different pipeline layers; here are their numbers against matched GRPO-family baselines.
Retrieval-Augmented Generation · vector retrieval vs agentic retrieval
Vector retrieval compresses a corpus into one similarity score per passage. Five papers measure the cost and how grep agents, small readers, citation grading, and per-chunk configs each recover a layer.
Code Generation · AlphaCode vs Code Llama
AlphaCode converts up to a million sampled candidates into a top-54.3% Codeforces finish; Code Llama ships open 7B-70B weights that hit 67% HumanEval: one is search, the other is infrastructure.
Reinforcement Learning · Self-distillation in RLVR vs Privileged-context teachers vs Dense token-level rewards
RLVR gives one bit per trajectory. SDPG copies a hint-conditioned teacher on correct trajectories, SDAR gates the copy per token, AntiSD rewards disagreement: three ways to mine dense signal from one privileged context.
AI Agents · SkillOpt vs Ctx2Skill vs OpenSkill vs LatentSkill vs SkillAdaptor vs SkillsVote
Six 2026 methods teach LLM agents skills without retraining: SkillOpt edits one skill document for +23.5 points, Ctx2Skill mines skills by self-play, OpenSkill learns unsupervised from the open web, plus three more.
AI Agents · RHO vs Harness-1
RHO tunes an existing harness from unlabeled trajectories; Harness-1 designs the harness so RL only learns semantic decisions. Same premise, different problems, each with its own numbers.
Code Generation · SWE-Explore vs DeNovoSWE vs Claw-SWE-Bench
Three 2026 papers isolate what moves coding-agent scores: localization stalls at 0.15-0.20 line recall, training lifts a 30B model 8x, and the harness alone swings Pass@1 by 27.4 points.
Speech Synthesis · VALL-E vs NaturalSpeech 2 vs SwanVoice
VALL-E turns speech into codec-token language modeling with a 3-second prompt, NaturalSpeech 2 diffuses continuous codec latents for prosody and singing, and SwanVoice generates a whole 1-4 speaker dialogue in one pass.
Segmentation · Mask2Former vs Mask R-CNN vs SAM vs SAM 2
Mask2Former beats Mask R-CNN 43.7 to 37.2 COCO AP at ResNet-50 in one table; SAM gets 46.5 zero-shot COCO AP vs 51.0 supervised ViTDet-H; SAM 2 hits 76.8 J&F on SA-V val vs XMem 60.1, 6x faster than SAM.
Text Embeddings · Sentence-BERT vs SimCSE vs E5
Sentence-BERT cut 10k-sentence similarity from 65 hours to 5 seconds; SimCSE turned dropout into free training pairs for 76.25 unsupervised and 81.57 supervised STS; E5 beat BM25 zero-shot across 56 retrieval datasets.
Agent Memory · MemGPT vs MRAgent vs δ-mem vs EvoMem
MemGPT pages memory like an OS, MRAgent rebuilds a memory graph while reasoning, δ-mem compresses history into an 8×8 state, and EvoMem stores change patches: four answers to where agent memory should live.
Self-Supervised Learning · SimCLR vs BYOL vs MAE vs DINOv2
SimCLR hit 69.3 ImageNet linear with negatives and 4096 batches; BYOL dropped negatives for 74.3 and small-batch robustness; MAE owns fine-tuning at 87.8 but freezes poorly; DINOv2's frozen 86.3 beats OpenCLIP 86.2.
AI Theorem Proving Papers · AI theorem provers vs formal mathematics
AlphaGeometry, DeepSeek-Prover-V1.5 and MaxProof all report headline math scores, but they prove different kinds of theorems under different verifiers and search budgets. Here is what each number actually measures.
Video Generation · Stream-T1 vs Echo-Infinity vs SANA-Streaming vs Causal Forcing++
Stream-T1, Echo-Infinity, SANA-Streaming and Causal Forcing++ all claim real-time video, but solve four different problems, and their FPS numbers are not comparable.
Vision-Language-Action · RT-2 vs π0 vs Qwen-VLA vs MolmoAct2 vs World Pilot
RT-2 wrote actions as text tokens, π0 switched to ~50Hz flow-matching chunks, and 2026 models add 3D reasoning or world-model priors. Simulation is nearly saturated; real-robot margins and OOD shifts differentiate.
Small Language Models · Phi-3 vs SmolLM2 vs MobileLLM vs TinyLlama
Below 4B parameters there is no winner: phi-3-mini hits 69 MMLU on curated data, SmolLM2 beats Llama3.2-1B on general benchmarks but trails Qwen2.5-1.5B on math, and MobileLLM adds 2.7 to 4.3 percent at 125M to 350M.
Reinforcement Learning · DelTA vs AntiSD vs DVAO
RLVR grades a whole reasoning trace with one bit at the end. DelTA reweights the token update, AntiSD adds a per-token PMI reward, DVAO rebalances several rewards by group variance — three fixes at three layers.
Alignment · RLHF vs DPO
RLHF trains a reward model and runs PPO on top of it; DPO deletes both with a closed-form rewrite and a single classification-style loss. The numbers say 'as good or better, far simpler', with real caveats.
Diffusion Language Models · Diffusion language models vs Autoregressive language models
An 8B masked-diffusion LLM matches an 8B autoregressive one on MMLU, beats it on GSM8K with a sixth of the tokens, and fixes the reversal curse. Decoding speed and serving maturity still favor autoregression.
Reinforcement Learning · GRPO vs PPO
GRPO swaps PPO's learned value model for the mean reward of a sampled group. That frees one policy-sized model and made open reasoning RL affordable, at the cost of per-token credit assignment.
Sequence Modeling · Mamba vs Transformer
Mamba swaps attention's quadratic compute and growing KV cache for a fixed-size selective state: 5x decode throughput and parity with 2x-larger Transformers at 3B, but exact recall stays with attention.
Fine-Tuning & Adaptation · LoRA vs Full fine-tuning
LoRA matches full fine-tuning within a point on standard tasks at 10,000x fewer trainable parameters, but it gets there through intruder dimensions that carry its forgetting and hurt under continual training.
Language Models · Muon vs AdamW
Muon reaches AdamW's loss with about half the pretraining FLOPs and now trains DeepSeek-V4, Kimi K3 and Qwen3.8-Next; used for SFT of an AdamW-pretrained model it loses points, and in RL it collapses.
Long Context · Sparse attention vs Full attention
Sparse attention buys 7x to 28x less attention compute at 1M tokens for roughly half a point on long-context suites, but the wall-clock gain is far smaller than the FLOPs gain and most of it is invisible at 32K.
World Models · World models vs LLMs
An LLM reasons over rules and goals in language; a world model predicts what a scene does next. The 2026 evidence says they are complementary, and the hard part is knowing when a generated future is credible.