RSD
Process rewards route reasoning steps between small and large models. FLOPs savings are promising; service-level latency remains unmeasured.
Read report
Blog posts
| Technology area | Oct 2026 | Sep 2026 | Aug 2026 | Jul 2026 |
|---|---|---|---|---|
| Agent Memory & Security | — | — | ||
| VLM, Vision & 3D | ||||
| LLM Systems & Efficient Inference | ||||
| Reinforcement Learning & PEFT | — | |||
| Trustworthy ML & Learning Theory | — | |||
| AI4Science & Scientific ML | — |
October 2026
LRAM
Modern recurrent backbones for large action models: long-context efficiency, action-copy shortcuts, and limits of simulation-only evidence.
Read reportTransition Matching
Learnable stochastic transitions connect flow matching and continuous autoregression, trading expensive backbone calls for small generative heads.
Read reportAIM-CoT
Information gain selects visual evidence while attention changes determine when to insert it into multimodal reasoning.
Read reportReActR
Error-type curricula and staged rewards teach mathematical reflection, with open questions about calibration and reproducibility.
Read reportPallatom
All-atom protein co-design through coordinate diffusion and recycling, balancing sequence quality, structural diversity, and experimental validation.
Read reportAmongUs
Social-deduction games reveal how deception and detection scale, and what internal probes can tell us about agent behavior.
Read reportSuperInfer
Efficient bidirectional KV migration for disaggregated inference, with a close look at latency targets and hardware dependence.
Read reportPointMapPolicy
Structured point maps connect visual geometry to robotic action, exposing task-dependent tradeoffs between spatial and color information.
Read reportSeptember 2026
DFM
Multiscale flow matching for image generation: interface simplicity, high-frequency artifacts, and reproducibility tradeoffs.
Read reportSpatial IDs
Low-rank object–position bindings in vision-language models, examined through interventions and spatial-reasoning tests.
Read reportRGM
Compact process reward models guide test-time reasoning, raising questions about full search costs and error-detection quality.
Read reportI2Mole
Discrete environment representations for drug-interaction prediction under distribution shifts.
Read reportCircuit Insights
WeightLens produces clearer circuit descriptions, while purity and causal faithfulness expose limits of interpretability metrics.
Read reportSeesaw
Switching tensor and pipeline parallelism between prefill and decode improves throughput while trading off latency and memory.
Read reportBridgeVLA
Three-view visual heatmaps connect pretrained vision-language representations to 3D robotic action keypoints.
Read reportHPS
Hard rejected-response sampling improves preference optimization, with careful attention to reward margins and external quality.
Read reportMolJO
Joint gradient guidance for molecular optimization, separating docking-score improvements from evidence of drug quality.
Read reportTilted Empirical Risk
Negative tilting for robust learning: theoretical guarantees, contamination models, and the challenge of choosing parameters in practice.
Read reportAda-KV
Adaptive cache budgets across attention heads for efficient long-context inference, with safeguards and deployment caveats.
Read reportEmbodied-R1
Geometrically verifiable reinforcement learning for embodied reasoning and action, connecting visual evidence to robotic control.
Read reportMeanFlow
Average-velocity learning for one-step generation, examining guidance, numerical consistency, and quality–compute tradeoffs.
Read reportMM-RLHF
Multimodal preference learning and safety alignment, with a review of dataset consistency and external evaluation.
Read reportThinkSLM
How quantization and structured pruning affect reasoning in small language models, beyond final-answer accuracy.
Read reportPemNN
Physical knowledge improves learning from scarce biophysical measurements, while input differences complicate attribution.
Read reportCoordRobustLR
Coordinate-wise corruption in linear regression: minimax limits and why even sparse missingness can be difficult.
Read reportScaleSearch
Scale selection for low-bit inference, disentangling quantization gains from high-precision safeguards.
Read reportVIKI-R
Reinforcement learning for collaborative embodied reasoning, with mixed in-distribution and out-of-distribution results.
Read reportAmbient Creativity
Noise-conditioned learning reduces diffusion-model memorization while exposing the gap between reduced copying and privacy guarantees.
Read reportDEFAME
Tool-assisted multimodal fact-checking, where evidence coverage matters more than simply extending reasoning.
Read reportATLANTIS
Training-data selection for instruction tuning, examining empirical gains and the assumptions behind importance weighting.
Read reportFMG
Formal molecular grammars guided by semantic feedback, balancing synthesizability, diversity, and limited training data.
Read reportEmergent Misalignment
Narrow fine-tuning can induce broader misalignment, with important differences across framing, models, and evaluation settings.
Read reportLumos
Trace-based performance modeling for distributed training, connecting runtime behavior to software–hardware planning.
Read reportDreamGen
Synthetic video as a source of executable robot-training signals, with tradeoffs in compute, transfer, and evidence quality.
Read reportLongLive
Streaming long-video generation: temporal consistency, persistent context, and efficient inference.
Read reportAVUT
Audio-visual understanding for video question answering, examining cross-modal evidence and evaluation.
Read reportAugust 2026
TIMING
Path attribution for time series: explaining temporal dependencies and evaluating the faithfulness of model explanations.
Read reportDocVXQA
Sufficiency, minimality, and token interactions for document-QA explanations, with a focus on faithfulness.
Read reportQServe
W4A8KV4 quantization and GPU kernel co-design remove dequantization bottlenecks to improve inference throughput.
Read reportTARL
Reward-free test-time policy adaptation through action-entropy minimization and KL constraints under environment shifts.
Read reportQ-VDiT
Token-aware low-rank compensation and video-relation distillation mitigate low-bit video Transformer degradation.
Read reportVGA
Visual Graph Arena tests graph recovery from pixels, layout robustness, and the separation of perception from reasoning.
Read reportARFT / AutoResearchEval
Process and artifact diagnostics across 800 scientific-agent trajectories reveal failures to revise detected flaws.
Read reportAlignment Potential
The teacher–student gap between reward models and current policies guides preference-data selection for alignment.
Read reportKLDM
Auxiliary velocities turn diffusion on periodic crystal coordinates into a dynamical system for crystal generation.
Read reportLIMEFLDL
A simplex-constrained attribution matrix explains the complete label distribution of structured-output models.
Read reportPARQ
Piecewise-affine regularization and proximal updates move quantization training from continuous to discrete weights.
Read reportSimbaV2
Hyperspherical normalization and scale-stable distributional value prediction improve deep RL scaling and stability.
Read reportDM/FM Alignment
Reward-weighted distributions unify diffusion and flow-matching alignment through changes to sampling dynamics.
Read reportMMedPO
Lesion-level perturbations and clinical relevance connect medical preference optimization to visual evidence.
Read reportHEX
Cross-block consensus enables test-time scaling in diffusion language models through implicit semi-autoregressive experts.
Read reportDPLM-2.1
Bit-level supervision, geometric biases, and representation alignment improve multimodal protein structure learning.
Read reportEARL-BO
Multistep query policies learned inside a Gaussian-process world model for nonmyopic Bayesian optimization.
Read reportTempo
A question-aware video compressor allocates a global token budget to improve efficient long-video understanding.
Read reportGeoRA
Gradient geometry guides trainable-parameter selection for low-rank adaptation under verifiable reinforcement learning rewards.
Read reportMolGraphBench
A systematic evaluation of molecular-regression GNNs reveals layer choice as a consequential hyperparameter.
Read reportTIS
Theoretical bounds on Transformer succinctness, architectural complexity advantages, and the conditions behind them.
Read reportSubSpec
A low-bit copy of the target model serves as a training-free draft for lossless speculative decoding with offloaded models.
Read reportOCR-Memory
Visual representations for long-term memory: optical context retrieval and the tradeoffs between capacity, cost, and fidelity.
Read reportExperience-Following
How historical experience creates feedback loops in agent behavior, and why more memory does not always help.
Read reportAgent-STAR
Long-horizon tool-use reinforcement learning through curriculum data, reward design, and generalization.
Read reportRnG
A unified Transformer uses implicit KV-cache representations to reconstruct 3D scenes from partial observations.
Read reportMAGMA
Four complementary relation graphs organize agent memory. Evidence and limitations from LoCoMo and LongMemEval.
Read reportNEMORI
Prediction error as a signal for adaptive memory distillation: deciding which experiences an agent should retain.
Read reportEMO-R3
Rewarding both reasoning structure and effectiveness in subjective visual tasks, with implications for RL evaluation.
Read reportETM
How reasoning traces can encourage overinterpretation, misplaced trust, and evaluation bias.
Read reportEvo 2
Long-context genomic modeling and unified biological foundation models for genome understanding and design.
Read reportGDPT
Grokking as a dimensional phase transition, connecting memorization, generalization, and experimental evidence.
Read reportJuly 2026
Graph of Tokens
Token relationships stabilize routing and load balancing in sparse mixture-of-experts models.
Read reportGAM
Event-progress graphs and thematic networks balance rapid adaptation with stable long-term knowledge.
Read reportFARMA
Forged reasoning memories expose a new attack surface in agent memory systems. A review of risks, evidence, and defenses.
Read reportINSID3
Training-free in-context segmentation: methods, quantitative evidence, and limitations from a CVPR 2026 Oral paper.
Read reportAgeMem
Unified long- and short-term memory management for language-model agents, with a close look at policy design and evaluation.
Read report