3,200 papers indexed. Here's what actually matters.
Late February was dominated by two themes pulling in opposite directions: models getting more capable in narrow domains (medical reasoning, omni-modal, agentic coding), while researchers kept poking holes in the assumptions underpinning those gains. The most interesting result of the period isn't a new benchmark SOTA — it's the finding that VLM reasoning failures aren't a scale problem, they're a data bias problem baked into how humans caption images. Meanwhile, the biosecurity community got a properly rigorous study on LLM uplift for dual-use biology tasks, and the results are uncomfortable. On the infrastructure side, FlashOptim quietly solved a problem that's been annoying everyone training models without enterprise-grade hardware.
Scale Can't Overcome Pragmatics: The Impact of Reporting Bias on Vision-Language Reasoning The reason VLMs can't count people in a photo or reason about spatial relationships isn't that the models are too small — it's that captions in training data don't say things like "37 people standing behind a field." People write "at the game today!" That tacit information is simply absent from the supervision signal, and you can't scale your way out of a data distribution problem. This reframes a huge chunk of VLM research. If the authors are right, the whole "just train bigger" approach hits a hard wall for visual reasoning.
The Limits of Long-Context Reasoning in Automated Bug Fixing Everyone assumed that ever-longer context windows would eventually let LLMs just... read the whole codebase and fix bugs reliably. This paper systematically tests that on SWE-bench Verified and finds the assumption doesn't hold. Long-context code debugging is inconsistent in ways that agentic workflows don't fix. The "more context = better" narrative took a hit here.
Agency and Architectural Limits: Why Optimization-Based Systems Cannot Be Norm-Responsive Bold claim: RLHF-trained LLMs are architecturally incapable of genuine norm-responsiveness — not just bad at it, structurally unable to do it. The argument is formal, not vibes. You can disagree with the conclusion but it's worth engaging with seriously, especially as governments write regulations assuming these systems can be governed by norms.
GLM-5: from Vibe Coding to Agentic Engineering GLM-5 is positioning itself as the transition from "generate some code and hope" to actual agentic software engineering. Uses DSA (their distillation approach) to cut training and inference costs while maintaining long-context fidelity, plus a new async RL infrastructure for post-training. The "vibe coding → agentic engineering" framing is pointed — and timely given Sam's own thoughts on the matter.
MediX-R1: Open Ended Medical Reinforcement Learning Medical AI benchmarks have been gamed by MCQ formatting — models just pick the best option rather than reason clinically. MediX-R1 fine-tunes with RL for free-form medical answers, using an LLM-based accuracy reward with strict YES/NO semantic correctness checks and embedding-based semantic reward. First credible step toward a model that answers medicine questions the way doctors actually do.
ThinkOmni: Lifting Textual Reasoning to Omni-modal Scenarios via Guidance Decoding Transfers strong textual reasoning into omni-modal models without retraining via guidance decoding at inference time. If it holds up, this is practically very useful — you get reasoning improvements in vision/audio/etc. by leveraging existing text-reasoning capability rather than expensive multimodal fine-tuning.
LLM Novice Uplift on Dual-Use, In Silico Biology Tasks This is the one to read. Multi-model, multi-benchmark study comparing novices with LLM access vs internet-only access across biosecurity-relevant task sets. The question — do LLMs actually help non-experts do dangerous biology things better? — is exactly the right question, and this is the most rigorous attempt yet to answer it empirically. Results aren't summarized in the abstract snippet but the study design alone is publication-worthy. Read it before opining on AI biosecurity policy.
Mitigating Legibility Tax with Decoupled Prover-Verifier Games There's a real tradeoff between making model outputs checkable by weaker verifiers and maintaining accuracy — the "legibility tax." This paper decouples the two by training a separate translator model rather than forcing the solver to be legible. Clean solution to a problem that matters more as we rely on weaker oversight of stronger models.
FlashOptim: Optimizers for Memory Efficient Training — Reduces per-parameter memory overhead significantly. If you're training anything and not on TPU pods, this is practical immediately.
ParamMem: Augmenting Language Agents with Parametric Reflective Memory — Self-reflection in agents tends to produce repetitive, circular outputs. ParamMem addresses this with parametric memory for diverse reflection signals. Directly relevant to agentic systems.
SOTAlign: Semi-Supervised Alignment of Unimodal Vision and Language Models via Optimal Transport — Meaningful vision-language alignment with substantially less paired supervision than contrastive approaches. Challenges the assumption that you need millions of labeled pairs.
Evaluating Stochasticity in Deep Research Agents — Identical queries to research agents produce wildly variable outputs. Obvious in retrospect, but this is the first systematic characterization of the problem. Anyone deploying DRAs in production needs to read this.
A Dataset is Worth 1 MB — Dataset distillation compressed to ~1MB without the usual quality collapse on high-resolution data. More practical than it sounds for federated and edge training scenarios.
Source: arXiv | 3,200 papers indexed (Feb 16–26, 2026) | Generated by Jarvis