This year's highest-rated papers
A benchmark of 390 AI papers measures scientific slop across structure, argument, and artifacts; a harness reduces the AI-human gap by 63% via evidence-grounded revision.
Published Sep 30, 2026 · 0 citations · ▲ 55 on Hugging Face · Code ★ 13
Only vote on papers you've read. Sign in with GitHub to vote.
MotorMind connects general vision-language models to deterministic robot control via mid-level actions and feedback loops, achieving 66.7% zero-shot success on LIBERO-PRO and 95% on real robots without task-specific training or external tools.
Published Sep 29, 2026 · 0 citations · ▲ 90 on Hugging Face
Only vote on papers you've read. Sign in with GitHub to vote.
DEPICT scores text-to-image alignment via expected agreement between image and caption answers, boosting negation accuracy to 88% and exceeding fine-tuned evaluators on human correlation benchmarks.
Published Oct 2, 2026 · 0 citations · ▲ 3 on Hugging Face
Only vote on papers you've read. Sign in with GitHub to vote.
Post-hoc OOD detectors inflate reported gains by fitting dataset identity rather than novelty, leaving only a single constant fit valid.
Published Oct 1, 2026 · 0 citations
Only vote on papers you've read. Sign in with GitHub to vote.
Post-training on protein-folding data via discrete answers and continuous geometry improves structure prediction and broad reasoning across ten benchmarks.
Published Sep 30, 2026 · 0 citations · ▲ 120 on Hugging Face · Code ★ 28
Only vote on papers you've read. Sign in with GitHub to vote.
OSWorld-Pro introduces process-based evaluation with 2,800 subgoals across 300 tasks, revealing top models achieve only 75.7% subgoal success versus 83.4% end-state performance and identifying distinct failure modes like irrelevant actions and click errors.
Published Sep 21, 2026 · 0 citations · ▲ 15 on Hugging Face
Only vote on papers you've read. Sign in with GitHub to vote.
OnePO enables RL-only medical domain adaptation via adaptive objective evolution and teacher retirement, yielding HuatuoGPT-3 that surpasses frontier models.
Published Oct 5, 2026 · ▲ 15 on Hugging Face · Code ★ 12
Only vote on papers you've read. Sign in with GitHub to vote.
StateM improves long-horizon agent accuracy via durable-state harness scaling without model changes, reaching 95.3% on Terminal-Bench 2.1 and cutting API costs to about $15 versus $574.68.
Published Aug 15, 2026 · 0 citations · ▲ 452 on Hugging Face · Code ★ 1,312
Only vote on papers you've read. Sign in with GitHub to vote.
DeskForge generates dense desktop supervision via controllable real-app environments, yielding 1.2M observations that improve GUI grounding and long-horizon computer-use task completion.
Published Oct 1, 2026 · 0 citations · ▲ 10 on Hugging Face · Code ★ 4
Only vote on papers you've read. Sign in with GitHub to vote.
UndoBench separates tool-using agent competence from fault recovery via paired enterprise workflow trials, finding 83.54% nominal success but only 46.72% recovery success with phase-dependent vulnerabilities.
Published Oct 4, 2026 · ▲ 11 on Hugging Face · Code ★ 1
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Post-training sharpens base model behaviors at the cost of solution coverage, introducing a quantifiable "Sharpening Tax"; a posterior-tempered group sampler reduces this tax while boosting accuracy.
Published Oct 1, 2026 · 0 citations · ▲ 102 on Hugging Face · Code ★ 25
Only vote on papers you've read. Sign in with GitHub to vote.
CurveCodec 2 predicts quantized skeletal curves from past values and learns residual entropy models to reduce animation storage to 0.22-0.37x of ACL with verified error bounds.
Published Oct 3, 2026 · ▲ 4 on Hugging Face · Code ★ 5
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
QuantWM is a training-free 2-bit KV cache quantization framework for video world models that preserves attention logits and token selection to eliminate temporal flickering while achieving up to 6.20x memory compression.
Published Sep 22, 2026 · 0 citations · ▲ 13 on Hugging Face
Only vote on papers you've read. Sign in with GitHub to vote.
The RA-Bench benchmark reveals current detectors fail to consistently identify AI-generated crisis videos, which become harder to detect after social dissemination and frequently mislead humans.
Published Aug 14, 2026 · 0 citations · ▲ 287 on Hugging Face · Code ★ 132
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
LoHi trades per-frame resolution for denser temporal sampling via low-resolution streams plus sparse high-resolution frames, boosting long-video accuracy up to 10.6 points and cutting front-end latency up to 7x.
Atlanta Poster Session 1, Wed, Dec 9, 10:00 AM–1:00 PM, Hall C1 · Published Oct 3, 2026 · ▲ 13 on Hugging Face
Only vote on papers you've read. Sign in with GitHub to vote.
Self-evolving search agents suffer co-cheating where proposers and solvers mutually reinforce errors; CrossFit partitions sources to cross-fit agreement and cuts false agreement by over half, boosting downstream search by 8+ points.
Published Sep 30, 2026 · 0 citations · ▲ 670 on Hugging Face
Only vote on papers you've read. Sign in with GitHub to vote.
LSP learns low-rank subspaces end-to-end via joint orthogonal projector optimization to reduce transformer memory and compute while outperforming local criteria at high compression ratios.
Published Sep 30, 2026 · 0 citations · ▲ 1 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
OpenART scales agent red teaming via open-ended environment evolution across 10,000 stateful scenarios, with EMHA achieving 85% attack success that grows with complexity.
Published Aug 1, 2026 · 0 citations · ▲ 266 on Hugging Face · Code ★ 231
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Automatic taxonomies reveal retrieval and verifier reasoning failures persist across scaled medical retrieve-then-verify systems, showing fundamental open-ended evaluation limits.
Published Sep 24, 2026 · 0 citations · ▲ 12 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Tool-using agents often fail by acting before establishing required evidence or leaving multi-action workflow prerequisites unresolved, despite accurate static action assessment. SafeActBench reveals failures stem from how agents use established evidence during execution, not just missing informatio
Published Oct 6, 2026 · ▲ 20 on Hugging Face · Code ★ 3
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
WING distills interaction-centric latent actions from egocentric video and uses cross-embodiment spectral low-frequency guidance to transfer them to robot policies, achieving high success rates on LIBERO, RoboTwin, RoboCasa, and real-world tasks.
Published Oct 2, 2026 · 0 citations · ▲ 20 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
World Models' Last Exam in Physics benchmarks video models via 40 measurement-based physics tasks, finding the best model scores 57.76/100 with widespread inconsistencies.
Published Oct 6, 2026 · ▲ 2 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Rhetorical robustness requires stable judgments across content-preserving rewrites and discrimination across papers; SciCore improves both via dual-branch science-core review.
Published Sep 30, 2026 · 0 citations · ▲ 76 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
LMBuild evaluates LLM agents on generating buildable, functional 3D structures and finds physical operability and functional affordance remain challenging despite improved soundness.
Published Oct 3, 2026 · ▲ 24 on Hugging Face · Code ★ 1
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
PDE-JEPA introduces predictive masked-latent pretraining with geometry projection and structured latent predictors for parametric PDE dynamics, reducing errors by 33.4% in-distribution and 51.4% on unseen parameters.
Published Sep 28, 2026 · 0 citations · ▲ 53 on Hugging Face · Code ★ 9
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
ProWAM predicts sparse visual sub-goals and actions via progressive planning, achieving state-of-the-art long-horizon robotic control and strong zero-shot real-world generalization.
Published Oct 1, 2026 · 0 citations · ▲ 83 on Hugging Face
Only vote on papers you've read. Sign in with GitHub to vote.
Self-generated feedback in long-horizon test-time training causes weight updates that improve synthetic text but degrade real-text prediction, and settlement on independent evidence prevents this failure.
Published Oct 4, 2026 · ▲ 19 on Hugging Face · Code ★ 1
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
AGO is an evidence-first quality gate for RAG that treats judge errors and missing data as explicit outcomes, using stratified beta-binomial gates and mandatory meta-evaluation to reduce unsafe promotion to 22.2%-35.1% versus 29.3%-41.8% for naive gates.
Published Oct 1, 2026 · 0 citations · ▲ 13 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
JLD defines a perceptual image distance via a Jacobian-derived metric tensor from frozen vision encoders, achieving state-of-the-art correlation with human judgments and resolution robustness.
Published Oct 5, 2026 · ▲ 3 on Hugging Face · Code ★ 3
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
LLM novelty judges are unstable: small prompt changes alter verdicts on over half of identical idea pairs and shift accuracy by over 50 points, undermining automated ideation evaluations.
Published Oct 1, 2026 · 0 citations
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Multilingual GSM-Symbolic introduces matched math problems across 15 languages to show model size, resource level, reasoning, and typology determine cross-lingual transfer, with size and reasoning closing low-resource gaps but not typological ones.
Published Oct 2, 2026 · 0 citations · ▲ 47 on Hugging Face · Code ★ 4
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
PointWAM forecasts 3D point trajectories of scenes and hands in a shared space-time frame to guide dexterous robot manipulation, improving DexJoCo success by 56.9 points with video pre-training.
Published Oct 2, 2026 · 0 citations · ▲ 42 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
MM-ABC is a mobile manipulation foundation model combining multi-level vision features, future imagination supervision, and masked joint attention to coordinate arm-base actions, achieving up to 99.1% success across benchmarks and 83% in real-world tasks.
Published Sep 28, 2026 · 0 citations · ▲ 5 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Deepfake detectors degrade to 76% accuracy and near-zero under attacks, so calibrated reconstruction-based authentication bounds false real-image certification to 1%.
Published Oct 5, 2026 · ▲ 13 on Hugging Face · Code ★ 2
Only vote on papers you've read. Sign in with GitHub to vote.
Existing token-reweighting methods cannot reverse harmful SFT features; SCALE uses frozen SFT deltas with entropy-guided gates to suppress, reverse, or extrapolate them, improving math and code results.
Published Sep 27, 2026 · 0 citations · ▲ 11 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
RRSI regularizes recursive agent harness self-improvement via annealed edit budgets, trajectory exploration, and critical selection to boost out-of-distribution performance and reduce token use. It improves up to 14.1 points in-distribution and 4.7 points out-of-distribution while cutting policy tok
Published Sep 21, 2026 · 0 citations · ▲ 222 on Hugging Face · Code ★ 1,274
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
EyeRobot 2.0 uses active gaze and fixation-relative frames to enable precise bimanual manipulation with only a single stereo camera, outperforming passive stereo by 40% in real-world trials and doubling ego-plus-wrist success under occlusion.
Published Oct 2, 2026 · 0 citations · ▲ 10 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Embodied progress reward models fail at long tasks due to missing context, but ProgressCompass supplies needed context to cut progress estimation errors by up to 82%.
Published Sep 29, 2026 · 0 citations · ▲ 11 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
OPUS defines optimizer-induced update-space data utility for dynamic LLM pre-training selection, outperforming full-scale baselines with minimal overhead.
Published Feb 5, 2026 · 0 citations · ▲ 354 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
MolmoAct2 is an open vision-language-action model with a specialized reasoning backbone, open action tokenizer, continuous-action expert, and adaptive reasoning that outperforms closed and open baselines across embodied reasoning and robot deployment benchmarks.
Published May 4, 2026 · 0 citations · ▲ 357 on Hugging Face · Code ★ 794
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
LLM scorers with equal ranking quality make unstable threshold and preference decisions under candidate reordering, and order-consistency fine-tuning fixes it without harming quality.
Published Aug 27, 2026 · 0 citations · ▲ 17 on Hugging Face · Code ★ 2
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
On-policy power distillation trains models to generate sharpened answers directly, improving single-sample math reasoning by up to 27.3 points and outperforming multi-candidate sampling and reward-based methods.
Published Oct 5, 2026 · ▲ 4 on Hugging Face · Code ★ 1
Only vote on papers you've read. Sign in with GitHub to vote.
DAEDALUS bootstraps reusable agent memory from self-generated practice tasks without oracles, improving success rates by up to 15.9 points across benchmarks.
Published Oct 6, 2026 · ▲ 5 on Hugging Face · Code
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
QuantCode specializes LLMs for executable trading code via framework pretraining and validated fine-tuning, boosting backtest success to 83.5% while revealing specialization trade-offs in tool use and repair.
Published Sep 30, 2026 · ▲ 11 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Typed decision models exhibit option-label bias because prompts prepend labels to definitions, letting label semantics override rules; removing labels or altering formatting fixes it.
Published Oct 1, 2026 · 0 citations · ▲ 5 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
HLA-WM combines geometry-guided retrieval with recurrent linear attention to fix long-range forgetting in video world models, improving 60-second consistency metrics by up to 28.5% with 12× lower memory and no retraining.
Published Oct 5, 2026 · ▲ 6 on Hugging Face · Code ★ 2
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
GeoSET is a generalist SAR-to-EO translation model pretrained on 3 million diverse pairs and adapted via LoRA, achieving state-of-the-art results across six benchmarks.
Published Sep 26, 2026 · 0 citations · ▲ 6 on Hugging Face · Code ★ 3
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
CARM prevents opposing token-level probability changes from canceling in sequence-level masking by using absolute log-ratios, improving RL reasoning and code benchmarks over geometric-mean masking.
Published Oct 1, 2026 · 0 citations
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
SlimWise decouples MoE expert pruning across prefill and decode phases to boost serving throughput without sacrificing accuracy via direct KV cache reuse and selective distillation.
Published Sep 28, 2026 · 0 citations · ▲ 5 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
FrameMorrow guides historical frame selection via prospective tokens representing future needs, improving consistency and quality across diverse long-horizon video generators.
Published Sep 30, 2026 · 0 citations · ▲ 92 on Hugging Face · Code ★ 26
Only vote on papers you've read. Sign in with GitHub to vote.
CtrlCache accelerates interactive video world models via control-aware caching that detects action changes to reuse transformer residuals and apply frequency-mixed history guidance, achieving up to 1.41x speedups with improved quality.
Published Oct 6, 2026 · ▲ 1 on Hugging Face · Code
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Self-compensating VLA adapts online to robot execution errors via residual feedback, improving success over 30 points on physical arms and outperforming training-time robustness methods on RoboStress.
Published Sep 29, 2026 · 0 citations · ▲ 18 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Periscope arranges text chunks in a grid to build an evidence map via local and strided probes, letting frozen language models answer questions across multi-million-token contexts with sublinear cost and small GPU memory.
Published Oct 2, 2026 · ▲ 8 on Hugging Face · Code ★ 1
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
OneStreamer unifies streaming video perception, memory, and proactive response via shared generation, achieving top results on eight benchmarks with a 4B model.
Published Oct 1, 2026 · 0 citations · ▲ 229 on Hugging Face · Code ★ 157
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
GraphForge synthesizes workspace tasks and verifiers over real file evidence graphs to train working agents, and fine-tuning Qwen3.6-27B improves GDPVal, Workspace-Bench-Lite, and SpreadsheetBench II results.
Published Sep 30, 2026 · 0 citations · ▲ 146 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
DrillBench benchmarks autoregressive drillhole modeling, finding lithology-sequence models transfer more robustly than spatial methods, and combining pretraining with retrieval improves cross-province generalization.
Published Oct 1, 2026 · 0 citations
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
StudentSim trains LLM student simulators via pooled training and per-student specialization, outperforming GPT-5.4 on behavioral fidelity and guidance responsiveness across chess, writing, and math.
Published Sep 1, 2026 · 0 citations · ▲ 495 on Hugging Face · Code ★ 53
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
RACE predicts subskill transition timing to extend VLA action chunks reliably, reducing stop-and-go idle time ~5x on real robots while improving success rates.
Published Oct 5, 2026 · ▲ 13 on Hugging Face · Code ★ 3
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
UnAct uses gradient-free targeted activation interventions to unlearn model classes from few forget images without gradients, labels, or retained data, matching or exceeding SSD and LFSSD accuracy across datasets and preventing network collapse with scarce data.
Published Oct 3, 2026 · ▲ 4 on Hugging Face · Code ★ 1
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Software-in-the-loop reconstruction scales terminal-agent training by deriving verified tasks from existing scientific workflows without manual references, improving Terminal-Bench 2 performance to 53.56%.
Published Oct 2, 2026 · 0 citations · ▲ 10 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
SHIFT predicts harness utility via a local LLM to search multi-agent structures per query, achieving ~80% mean accuracy across benchmarks while reducing execution tokens by 32%.
Published Oct 2, 2026 · ▲ 10 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Video2Skill benchmarks streaming embodied skill discovery, showing VLMs group manipulation events poorly and rarely expand skill libraries despite supervised fine-tuning.
Published Sep 29, 2026 · 0 citations · ▲ 11 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
YuE2 unifies symbolic and audio music generation through symbolic planning, producing readable scores and full-song audio that outperform public baselines and rival proprietary generators.
Published Sep 27, 2026 · 0 citations · ▲ 245 on Hugging Face · Code ★ 10,871
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Targeted bias injection via closed-loop activation steering exploits diffusion language model denoising trajectories to steer frozen models toward adversarial demographic answers with minimal corruption.
Published Oct 5, 2026 · ▲ 14 on Hugging Face · Code ★ 3
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
GDPO decouples per-reward normalization in multi-reward RL to prevent advantage collapse, improving training stability and outperforming GRPO on reasoning and coding tasks.
Published Jan 8, 2026 · 0 citations · ▲ 235 on Hugging Face · Code ★ 512
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Accent analogy guidance subtracts estimated accent directions for cross-lingual voice cloning, raising speaker similarity above identity-accent trade-off curves across several open TTS models.
Published Sep 24, 2026 · 0 citations · ▲ 4 on Hugging Face · Code
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
ALoDLM applies token-adaptive latent recurrence to diffusion language models, allocating computation by difficulty to close the quality gap with autoregressive models at 1.7B and 8B scales.
Published Oct 3, 2026 · ▲ 55 on Hugging Face · Code ★ 4
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
EnvHarness wraps static environments with programmable components to reshape agent behavior without altering underlying logic, improving benchmarks by up to 9.0 points while enabling continuous policy-environment co-evolution.
Published Aug 20, 2026 · 0 citations · ▲ 175 on Hugging Face · Code ★ 618
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Human audit of 165 WebArena-Lite tasks recovers 5.45, 8.49% evaluator-missed successes, reveals trajectory errors like looping, and shows guide text and MASM improve results.
Published Oct 1, 2026 · 0 citations
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Agent Priors-guided Policy Learning embeds structural priors in skill interfaces to enable compositional and out-of-distribution skill generalization.
Published Sep 28, 2026 · 0 citations · ▲ 84 on Hugging Face · Code ★ 3
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Edge0 predicts next-layer MoE routing one token ahead to stream experts from SSD, serving 35B-class MoEs at 20 tok/s within 3 GiB active memory on a 24 GB machine via recovery LoRA adapters.
Published Sep 16, 2026 · 0 citations · ▲ 25 on Hugging Face · Code ★ 3,293
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
In controlled strong-to-weak distillation, rollout policy is less central than token-level KL direction and learning rate, though on-policy data can improve generalization on harder reasoning tasks.
Published Sep 28, 2026 · 0 citations · ▲ 194 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Context language models treat context as self-modified files to learn context management, outperforming external strategies with lower compute and enabling in-context and parametric learning of management strategies.
Published Sep 29, 2026 · 0 citations · ▲ 40 on Hugging Face · Code ★ 571
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
VeriHarness turns fixed base LLMs into agentic verifiers with workspaces and evidence tools, achieving top selection scores and 6.2, 6.4 point gains over single rollouts on long-horizon tasks.
Published Oct 1, 2026 · 0 citations · ▲ 49 on Hugging Face · Code ★ 42
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Proactive LLM agents need joint optimization of task capability, temporal compute allocation, and user trust, with Proactivity-Gym exposing evaluation gaps and human preference for unobtrusive assistance.
Published Sep 29, 2026 · 0 citations · ▲ 22 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
FailBank turns runtime shield feedback into persistent VLA policy updates via failure-bank self-evolution, raising success rates up to 25.4 points and cutting policy-induced cost up to 35.6%.
Published Sep 30, 2026 · 0 citations · ▲ 15 on Hugging Face · Code ★ 2
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
KeyRec creates bounded visual memory via recent caches and structured event banks to enable efficient long-video and streaming understanding with only 10% of visual tokens, outperforming compressed baselines.
Published Sep 26, 2026 · 0 citations · ▲ 11 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Direct-decision JEV models show ordinal scale-utilization bias, compressing decisions to 26, 76% of gold support despite high accuracy, but BA-LoRA post-training improves utilization to 86%.
Published Sep 30, 2026 · 0 citations · ▲ 60 on Hugging Face · Code ★ 3
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Queen, a 4B-parameter chess-language model, plays at grandmaster level and explains moves via cross-attention to a silent expert encoder and iterative Bellman-style explanation distillation, surpassing larger frontier models.
Published Oct 2, 2026 · 0 citations · ▲ 31 on Hugging Face · Code ★ 18
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
MetaRubric fixes vacuous rubric credit via evidence-aware optimization and counterfactual rubric adaptation, improving PubMedQA accuracy by up to 20.40 points over static-judge GRPO.
Published Oct 2, 2026 · 0 citations · ▲ 22 on Hugging Face · Code ★ 2
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
PaLoRA derives an optimal rank-aware pacing law for LoRA continual learning that adaptively restricts gradient scaling to prevent forgetting, improving long-horizon benchmark accuracy by 4%.
Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published Oct 3, 2026 · ▲ 9 on Hugging Face · Code ★ 2
Only vote on papers you've read. Sign in with GitHub to vote.
FairRSFM benchmarks remote sensing foundation models by biome to expose hidden ecological performance disparities and tests debiasing methods without backbone updates. Aggregate metrics consistently mask large biome-dependent gaps, though mitigation effectiveness varies by model and task.
Published Oct 5, 2026 · ▲ 5 on Hugging Face · Code
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Post-training multi-token prediction heads on ~2.5B chain-of-thought tokens match pretraining speedups with 10^3-10^4x less data, while chain-aware verification and adaptive head selection boost throughput up to 16%.
Published Oct 1, 2026 · 0 citations
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Mid-Harness verifies candidate terminal actions at the model-harness boundary, raising TerminalBench-Lite Pass@1 from 50.00% to 68.03% and improving success at lower token cost than trajectory scaling alone.
Published Sep 30, 2026 · 0 citations · ▲ 116 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Pretrained transformers stop following references after 1.4, 3.6 lines, but a rank-8 LoRA at one early layer extends computation to 50, 160 lines without changing frozen weights.
Published Sep 29, 2026 · 0 citations · ▲ 75 on Hugging Face · Code ★ 3
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Fixing initial token cues in base models boosts reasoning to match RL performance, with effects traced to training data associations that can be causally edited.
Published Oct 5, 2026 · ▲ 12 on Hugging Face · Code ★ 2
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Feasible-future decoding reranks VLA actions by future safe-completion mass, reducing cumulative safety costs by up to 57.5% without retraining or rollouts.
Published Oct 4, 2026 · ▲ 3 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
GUI-HARVEST optimizes executable harnesses for frozen GUI agents by aligning visual effects, comparing task runs, and consolidating failure patterns into reusable source edits, improving OSWorld-Verified by up to 12.33 points.
Published Oct 1, 2026 · 0 citations · ▲ 5 on Hugging Face · Code ★ 2
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Judgment models excel at evaluation but fail at simulation, yet pairing them with code simulation yields expert control.
Published Oct 1, 2026 · 0 citations
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
Cross-tokenizer on-policy distillation achieves comparable accuracy with strict top-16 shared-vocabulary supervision versus full coverage, while expanded span supervision reduces accuracy due to conflicting gradients, motivating prioritization of supervision reliability over alignment coverage.
Published Oct 6, 2026 · ▲ 40 on Hugging Face
Only vote on papers you've read. Sign in with GitHub to vote.
CARPM-FIQA accumulates relative point margin scores across training epochs to stabilize face image quality estimates, reducing variance and improving ranking stability near top performance.
Published Sep 15, 2026 · ▲ 6 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
EditHero benchmarks long-horizon part-level 3D editing via sequential instructions and target images, finding agentic code-based methods preserve unedited regions better but are slower than non-agentic regeneration.
Published Oct 1, 2026 · 0 citations · ▲ 50 on Hugging Face · Code ★ 12
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
AutoSciBench autonomously generates and iteratively adapts scientific agent benchmarks via recipes and concepts, reducing solver accuracy by over 22 points versus human benchmarks while improving quality ratings.
Published Oct 4, 2026 · ▲ 17 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
WEFT evolves whole agentic interaction systems for tool-use post-training, outperforming environment-scaling baselines by up to 12.27 points across benchmarks.
Published Sep 29, 2026 · 0 citations · ▲ 19 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
SimuVerity benchmarks text-to-executable Simulink generation across engineering domains, finding best agents score only 42.86 and structural similarity poorly predicts engineering performance.
Published Oct 1, 2026 · ▲ 47 on Hugging Face · Code ★ 19
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
FORESIGHT uses dual-stream anticipatory planning in frozen streaming VLMs to dynamically configure future perception, improving online benchmarks by up to 18.7 points without retraining.
Published Oct 2, 2026 · 0 citations · ▲ 3 on Hugging Face · Code
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
TextReg mitigates prompt distributional overfitting via regularized text-space optimization, improving out-of-distribution accuracy by up to 16.5% over prior methods.
Published May 20, 2026 · 0 citations · ▲ 7 on Hugging Face · Code
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
SkillOpt treats agent skills as external state optimized via bounded text edits validated on held-out scores, improving accuracy up to 24.8 points with stable transfer.
Published May 22, 2026 · 2 citations · ▲ 266 on Hugging Face · Code ★ 18,062
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
UNREAL unifies retrieval and long-context evidence selection via frozen LLM representations with minimal parameters, outperforming state-of-the-art retrievers and improving long-context accuracy substantially.
Published Oct 6, 2026 · ▲ 15 on Hugging Face
Readers and the AI panel: vote on this paper to see what they said.
Only vote on papers you've read. Sign in with GitHub to vote.
RecursiveMAS scales multi-agent collaboration through recursive latent-space computation via RecursiveLink and inner-outer loop co-optimization, improving accuracy by 8.3% with 1.2-2.4x speedup and 34.6%-75.6% token reduction over baselines.
Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published Apr 28, 2026 · 0 citations · ▲ 239 on Hugging Face · Code ★ 961
Only vote on papers you've read. Sign in with GitHub to vote.