• Mar 1, 2026
  • 01Latent reasoning gains come from side effects, not reasoning itself. Causal mediation analysis reveals a causal disconnect between latent tokens and both inputs and outputs. A simple text-based "imagination" baseline outperforms complex latent-space methods.
  • 02Deep research agent cuts 70% of reasoning steps and gets more accurate. Parallel evidence gathering replaces serial reasoning chains. Search breadth beats reasoning depth.
  • 03"Test-driven error correction" from education psychology enters multimodal training. A diagnostic-reinforcement loop automatically identifies model weaknesses and generates targeted data. Eleven benchmarks improve without interfering with each other.
  • 04A framework for what makes a world model "qualified" arrives, but practical application is far off. Triple consistency (modality, spatial, temporal) provides a unified evaluation lens. The 184 HF upvotes reflect community anxiety more than a solved problem.
  • 05Multi-agent error propagation gets a plug-and-play firewall. Test-time pruning intercepts misinformation flow without retraining or changing topology. Average improvement: 6.3 percentage points.
  Filtered from 353 papers ↗
• Feb 28, 2026
  • 01Apple trains a tri-modal masked diffusion model from scratch. Systematic testing of scaling laws, modality mixing, and noise schedules makes this directly actionable for teams working on multi-modal diffusion. Masked diffusion is emerging as a real alternative to autoregressive generation.
  • 02Agentic RL training collapse now has a diagnostic framework. ARLArena decomposes policy gradients into four design dimensions and ablates each one to find instability sources. More effective than swapping algorithms blindly.
  • 03SkyReels-V4 generates video and audio together via dual-stream MMDiT. Text-to-video, inpainting, and editing all collapse into a single interface. Unified architectures are absorbing standalone modality pipelines.
  • 04Adding CoT reasoning to GUI agents actually hurts grounding. GUI-Libra identifies the root causes: action token dilution and incomplete step-level verification. Targeted fixes follow.
  • 05World models go multi-player and multi-viewpoint. Solaris achieves consistent multi-agent simulation in Minecraft. The automated data collection system may outlast the model itself.
  Filtered from 351 papers ↗
• Feb 27, 2026
  • 01TTT architectures are formally equivalent to linear attention. NVIDIA's proof unifies two independent research communities, sharply narrowing the design space for efficient sequence modeling.
  • 02The training data recipe for terminal agents is finally public. From seed task generation to skill composition and training strategy comparisons, the full dataset and model weights are open-sourced. An 8B model jumps from 2.5% to 13.0% accuracy.
  • 03RL-trained vision agents have a laziness problem, and now there's an engineering fix. Oversampling plus cumulative tool rewards effectively stops interaction collapse, keeping models from degenerating into single-turn QA.
  • 04Multi-modal retrieval's storage bottleneck gets a universal compression solution. Attention-guided clustering compresses document vectors to a fixed budget while preserving retrieval quality across text, image, and video.
  • 05Google's Aletheia completes a math proof challenge fully autonomously. But 10 problems is nowhere near enough to draw conclusions. The maturity of math reasoning benchmarks may be the bigger bottleneck.
  Filtered from 332 papers ↗
• Feb 26, 2026
  • 01A 20-person red team exposed 11 agent failure modes in a real deployment environment with persistent memory, email, Discord, and shell access. The most dangerous: agents reporting "task complete" while the underlying state is already broken.
  • 02Video reasoning finally gets a million-scale benchmark. VBVR replaces model-based judging with rule-based scoring, shifting video evaluation from "visual quality" to "spatiotemporal causal understanding." 404 HF upvotes suggest the community has been waiting.
  • 03A unified multimodal model runs on a phone. Mobile-O redesigns the fusion architecture with depthwise separable convolutions instead of distillation. 74% on GenEval, 6x faster than Show-O, 3 seconds per image on an iPhone.
  • 04Multi-model routing drops from query-level to step-level. SkillOrchestra replaces end-to-end RL with explicit skill modeling, cutting training cost 300–700x and eliminating routing collapse.
  • 05Policy collapse in RLVR training is more common than expected. Token-level entropy regularization changes wording, not reasoning paths. DSDR intervenes at both trajectory and token level, improving accuracy and pass@k.
  Filtered from 343 papers ↗
• Feb 25, 2026
  • 01An LLM builds an internal "world model" of kernel behavior to plan optimization paths. On complex kernels like MoE, it runs 14x faster than evolutionary search, turning operator tuning from random trial-and-error into guided exploration.
  • 02VLM token probabilities double as reward signals. Pretrained model logits encode task-progress information. Zero-shot correlation hits 0.947 across 130+ real robot tasks.
  • 03Agent memory evaluation has structural flaws. Benchmarks saturate, metrics disconnect from semantic utility, and swapping the backbone model flips conclusions. A checklist for teams building agent systems.
  • 04Inverse distillation moves from continuous to discrete domains for diffusion language models. Solving uniqueness and gradient stability unlocks 4–64x step compression.
  Filtered from 171 papers ↗
• Feb 24, 2026
  • 0174% of enterprise workflow tasks don't need inter-agent coordination. Monotonicity analysis provides a formal test: if merging sub-results can't make things worse, the task can run fully in parallel with zero orchestration overhead.
  • 02Multiple AI analysts examining the same data frequently reach contradictory conclusions. The disagreement isn't noise. It's structural. Prompt wording and model choice color the outcome before analysis even begins.
  • 03Multimodal models understand video content fine, but accuracy drops sharply on step ordering. TPRU uses RL fine-tuning on a 7B model to beat GPT-4o on temporal reasoning tasks. Accepted at ICLR.
  • 04Agent trajectory data now has a black-box watermarking scheme. Hidden hook actions embedded at decision points achieve 94.3 detection AUC without requiring access to model weights.
  Filtered from 149 papers ↗
• Feb 23, 2026
  • 01Weight folding outperforms pruning at most compression rates. ICLR 2026 work proves folding yields lower reconstruction error and validates across 1,000+ checkpoints.
  • 02Video generation models can now track your fingers. Joint-level hand control makes XR scenes interactive, not just watchable.
  • 03VR conversational agents finally know where you're standing. SARAH generates spatially aware full-body motion at 300 FPS for streaming VR deployment.
  Filtered from 163 papers ↗
• Feb 22, 2026
  • 01Latent Diffusion's Two-Step Training Collapses Into One. Aligning encoder output noise with the diffusion schedule yields a unified objective. FID 1.4 on ImageNet-512 with lower training FLOPs.
  • 02DiT Denoising Doesn't Need Fine-Grained Patches at Every Step. DDiT adjusts patch size by content complexity and denoising stage. 3.5x speedup, no quality loss, no retraining.
  • 03A Full Validation of MoE Best Practices at Scale. Arcee Trinity combines sigmoid routing, interleaved attention, and gated attention in a 400B-parameter model trained on 17T tokens with zero loss spikes.
  • 04Mamba-2 Gets More Accurate by Removing Components. Systematic ablation produces a simplified variant that nearly matches softmax attention while keeping linear complexity.
  Filtered from 250 papers ↗
• Feb 21, 2026
  • 01Agents Scored 80→90, but Failure Modes Barely Changed. Testing 14 models shows capability gains don't translate to reliability gains. Demo-to-production decisions should hinge on failure conditions, not average accuracy.
  • 02VLM + Sim RL Bypasses the Demonstration Data Bottleneck. HERO lets humanoid robots manipulate never-seen objects zero-shot, cutting end-effector tracking error by 3.2x.
  • 03Fast Weight Long-Context Bottleneck Is the Training Objective, Not the Architecture. Switching to next-sequence prediction with RL makes fixed-memory models competitive on long-context tasks for the first time.
  • 04Cold Start and Preference Drift, Solved in One Framework. Princeton's PAHF uses continual learning with dual feedback channels so agents keep up with shifting user preferences.
  Filtered from 286 papers ↗
• Feb 20, 2026
  • 01Zhipu's GLM-5 goes open-source, but core architecture claims await verification. DSA simultaneously promises lower cost, longer context, and better reasoning — three goals that typically conflict. Community reproduction will matter more than official benchmarks.
  • 02LLM agents systematically favor certain sources when filtering information. CMU's controlled experiments across 12 models show source preference sometimes overrides content relevance, and explicit "stay neutral" prompts don't fix it.
  • 03LoRA basis decomposition parameterizes visual transforms as a continuous space. NVIDIA's approach lets a single pair of example images specify arbitrary visual edits — no text prompt needed for effects that are hard to describe in words.
  • 04Dual-path memory retrieval achieves state-of-the-art on long-term memory benchmarks. Mnemis layers hierarchical graph traversal on top of vector recall, directly useful for building conversational systems with persistent memory.
  Filtered from 261 papers ↗
• Feb 19, 2026
  • 01Better pretraining makes quantization worse. Amazon finds activation outlier severity scales with pretraining duration. S2D applies spectral decay during training to fix the root cause, recovering up to 7% accuracy under W4A4.
  • 02Most fine-tuning data selection tricks are a waste of effort. Microsoft Research systematically disentangles the components and finds only gradient-based representations reliably predict downstream performance. With enough data, careful selection barely beats random.
  • 03Token-level policy gradients fundamentally mismatch reasoning granularity. MPO packs consecutive K tokens into semantic actions for policy gradient, aligning the optimization target with actual reasoning structure.
  • 04Airbnb reformulates geo-retrieval as extreme classification over 25 million grid cells. This drastically narrows the candidate set before ranking, tackling the heterogeneity problem inherent in two-sided marketplaces.
  Filtered from 288 papers ↗
• Feb 18, 2026
  • 01Binary tokens replace codebook indices. BitDance matches 1.4B-parameter models with 260M parameters, runs 8.7x faster, and hits 30x+ speedup at 1024 resolution.
  • 02RL training feedback too sparse for models to learn? ERL adds explicit reflection on failures before reinforcing successes, with gains up to 81% in complex environments.
  • 03Training data for search agents is expensive and hard to build. REDSearcher synthesizes complex tasks via graph topology and pairs them with a local simulation environment to slash RL iteration costs.
  • 04Inference-time compute still relies on high-temperature sampling to get lucky? STATe replaces random sampling with structured reasoning templates — more controllable and more interpretable.
  Filtered from 156 papers ↗
• Feb 17, 2026
  • 01Online RL finally works for web navigation agents. OpAgent hits 71.6% on WebArena — more than doubling every previous monolithic baseline.
  • 02Reward models don't have to judge forward. FLIP infers backward to the instruction, and a 7–9B model beats LLM-as-Judge by 79.6%.
  • 03RL isn't just for alignment anymore — it now trains embedding reasoning chains. Embed-RL teaches cross-modal retrieval to "think before matching."
  Filtered from 160 papers ↗
• Feb 16, 2026
  • 01A medical multimodal model now outperforms GPT-4o-class closed-source systems. MedXIAOHE chains entity-aware pretraining with RL-based reasoning to cover everything from rare diseases to long-form report generation.
  • 02Xiaomi open-sources a robot VLA model that runs real-time bimanual manipulation on a consumer GPU. The key is asynchronous execution baked into training, not just deployment.
  • 03Scientific tool use is agents' Achilles' heel. SciAgentGym stress-tests 1,780 domain tools — and an 8B fine-tuned model beats a 235B general-purpose one.
  • 04RL fine-tuning boosts VLM benchmark scores, but chain-of-thought faithfulness degrades — surfacing a hidden accuracy-vs-reliability trade-off.
  Filtered from 239 papers ↗
• Feb 15, 2026
  • 01Combine solved easy problems into new hard ones. Composition-RL turns wasted RLVR training samples into effective composite challenges, with consistent gains from 4B to 30B models.
  • 025B parameters doing the job of 80B. DeepGen 1.0 beats opponents 10x its size in both image generation and editing — code and weights fully open-sourced.
  • 03Students can surpass their teachers. ExOPD breaks the distillation performance ceiling through "reward extrapolation," and multi-domain expert knowledge can be merged back into a single small model.
  • 041M-token context on a single A6000D. MiniCPM-SALA's sparse + linear attention hybrid cuts long-context inference cost to a third.
  Filtered from 352 papers ↗
• Feb 14, 2026
  • 01196B parameters, only 11B active, and it matches GPT-5.2. Step 3.5 Flash uses MoE + RL to push agent efficiency to a new frontier — with open weights.
  • 02Coding agents can fix bugs, but can they build features? FeatureBench upgrades evaluation from single-PR fixes to end-to-end feature development. The best model passes just 11%.
  • 03Mistral releases Voxtral Realtime, a streaming speech recognition model. 480ms latency matching Whisper's offline transcription quality. Apache 2.0.
  • 04Long-context reasoning gets a "brake pedal." GRU-Mem uses gated mechanisms so agents know when to update memory and when to stop — up to 4x faster inference.
  Filtered from 407 papers ↗
• Feb 13, 2026
  • 01AI just solved real open problems in mathematics for the first time. Google DeepMind's Aletheia agent autonomously cracked 4 unsolved questions from the Erdos Conjectures database.
  • 02GUI world models don't need pixel-level prediction. Code2World turns next-state forecasting into code generation — an 8B model rivals GPT-5.
  • 03Learning control policies straight from video. VideoWorld 2 improves task success rate by up to 70% on real-world handcraft tasks, and the knowledge transfers to robot manipulation.
  • 04A training-free reasoning framework that switches cognitive modes per step. Chain of Mindset beats the strongest baseline by nearly 5% across six benchmarks.
  Filtered from 397 papers ↗
• Feb 12, 2026
  • 01Text diffusion models are no longer a proof of concept. LLaDA2.1's 100B model hits 892 TPS on code tasks and is the first dLLM to undergo large-scale RL training.
  • 02Open-source video+audio joint generation is here. MOVA generates visuals, dialogue, sound effects, and music in a single model.
  • 03GUI agents that actually work on real phones. UI-Venus-1.5 sets new SOTA on ScreenSpot-Pro and AndroidWorld across three model sizes from 2B to 30B.
  • 04When post-training saturates, teach from your own weak checkpoints. WMSS uses earlier model states to recover forgotten capabilities at zero inference cost.
  Filtered from 456 papers ↗
• Feb 11, 2026
  • 01Long-running agents suffer "context rot," but good context management can claw back most of the lost performance. LOCA-bench is the first benchmark to systematically test agent degradation under dynamic context growth.
  • 02Forcing LLMs to "think" before acting makes agents worse, not better. Thinking makes them introverted — shorter replies, less information shared — and downstream tasks collapse as a result.
  • 03FlashVID keeps 99% of video understanding performance with just 10% of visual tokens. The freed-up compute can push input frame counts up 10x at the same budget.
  • 04To attack LLM inference latency, target the scheduler, not the model. The Fill and Squeeze strategy inflates time-to-first-token by up to 280x.
  Filtered from 185 papers ↗
• Feb 10, 2026
  • 01LinkedIn got LLM search ranking into production. Multi-teacher distillation into a small model plus prefill-optimized serving yields 75x ranking throughput with no latency increase — one of the most detailed public accounts of LLM search deployment.
  • 02Multi-GPU inference is bottlenecked by synchronization. Parallel Track Transformer cuts cross-device syncs by 16x, delivering 15–30% lower time-to-first-token on vLLM and TensorRT-LLM.
  • 03Multi-turn conversations drift off track — the root cause is structural intent misalignment, not model weakness. A Mediator-Assistant architecture decouples intent understanding from task execution.
  • 04No fine-tuning, no training — just composing existing steering vectors adapts LLMs to new tasks. Steer2Adapt dynamically discovers optimal combinations from a few examples, averaging 8.2% improvement.
  Filtered from 187 papers ↗

View all archive →

FAQ

I'm not a researcher — will I understand this?

Absolutely. I write assuming you don't read papers. Each write-up leads with "why this matters" before any technical detail, and jargon is explained inline. Most readers are founders, product managers, and indie developers.

How is this different from AI news on Twitter or newsletters?

Most AI news is secondhand — summaries of summaries, where signal degrades with each retelling. I work directly from arXiv source papers, use a rule-based engine to filter 300+ papers daily, then write practitioner-focused analysis. Every write-up links to the original paper so you can verify.

How are papers selected?

Automated scoring plus human editorial. New AI/ML papers are fetched daily from arXiv and scored across 8 signal types — institutional origin, community traction, code availability, and more. The top 10-20 make each briefing. Full details on the Methodology page.

How do I keep up with daily updates?

The easiest way is RSS (works great with Feedly or any reader app). Or just bookmark this page — one visit per weekday, 3-5 minutes, and you're caught up.