February 2026 will be remembered as the month the field started questioning its own assumptions at scale. Three converging threads defined the period: a damning finding that reporting bias in training data β not model size β is the root cause of VLM reasoning failures; a theoretical breakthrough proving model-free general RL is finally achievable; and a wave of efficiency work (memory, tokens, compression) that signals the community is no longer willing to pay the compute bill indefinitely. The safety literature quietly produced two papers with genuine teeth β one on biosecurity uplift that policy-makers should be reading, and one arguing RLHF systems are architecturally incapable of norm-following. Meanwhile, 863 computer vision papers dropped in ten days, which tells you something about where the field's attention actually is.
"Scale Can't Overcome Pragmatics: The Impact of Reporting Bias on Vision-Language Reasoning" Kamath, Hessel, Chandu et al. β Feb 26, 2026
This is the paper that should make every VLM team stop and rethink their data pipeline. The thesis: VLMs fail at certain reasoning tasks not because they're too small, but because the training data is systematically biased by how humans caption images. People write "at the game today!" not "a photo of 37 people standing behind a field." That tacit information β the stuff humans consider too obvious to say β is exactly what's needed for grounded visual reasoning, and it's absent by construction from any web-crawled dataset. The authors investigate OpenCLIP and other major VLMs, showing this bias is pervasive and structural. The implication is brutal: you cannot scale your way out of this. If you want VLMs that reason correctly about visual scenes, you need fundamentally different data collection strategies, not more of the same. This will be cited in every VLM reasoning paper for the next two years.
π https://arxiv.org/abs/2602.23351v1
"FlashOptim: Optimizers for Memory Efficient Training" Gonzalez Ortiz, Gupta, Renard et al. β Feb 26, 2026
Training a 7B parameter model currently requires >100GB of accelerator memory just for parameters, gradients, and optimizer states. FlashOptim attacks this directly with a suite of optimizations that cuts per-parameter memory substantially. This matters because the current regime where only well-funded labs can fine-tune frontier models is a research bottleneck, not a feature. If FlashOptim delivers on its claims, expect it to land in every serious training stack within six months.
π https://arxiv.org/abs/2602.23349v1
"A Model-Free Universal AI" (AIQI) Kim & Lee β Feb 26, 2026
A genuine theoretical landmark, and probably the most underappreciated paper this month. AIQI is the first model-free agent proven to be asymptotically Ξ΅-optimal in general reinforcement learning β meaning it works across arbitrary computable environments without maintaining an explicit world model. Every prior optimal agent (including AIXI) was model-based. The proof uses universal induction over distributional action-value functions rather than policies or environments. Practically near-useless today, theoretically significant forever.
π https://arxiv.org/abs/2602.23242v1
"Xray-Visual Models: Scaling Vision Models on Industry Scale Data" Mishra, Lin, Wang et al. (Meta) β Feb 18, 2026
15 billion curated image-text pairs. 10 billion video-hashtag pairs. Facebook and Instagram as the data source. Meta quietly dropped a vision model paper that redefines what "industry scale" means, and the data curation pipeline β balancing, noise suppression, semantic diversity maximization β is arguably the actual contribution. The three-stage training pipeline will become a reference architecture. Whether academic labs can replicate any of this is a different question.
π https://arxiv.org/abs/2602.16918v1
"SOTAlign: Semi-Supervised Alignment of Unimodal Vision and Language Models via Optimal Transport" Roschmann, Krzakala, Mazelet et al. β Feb 26, 2026
The Platonic Representation Hypothesis β that vision and language models converge toward the same statistical model of the world β has been floating around as a hypothesis. SOTAlign weaponizes it: using optimal transport, they align frozen pretrained unimodal models with drastically less paired supervision than contrastive approaches require. If meaningful cross-modal alignment can be achieved semi-supervised, the cost of building VLMs drops significantly. Watch this one.
π https://arxiv.org/abs/2602.23353v1
"Agency and Architectural Limits: Why Optimization-Based Systems Cannot Be Norm-Responsive" Sarma β Feb 26, 2026
A provocative formal argument: RLHF systems are architecturally incapable of genuine norm-following because they lack the two necessary conditions for agency β stable preference representations and the capacity to represent obligations as obligations rather than as reward signals. The claim is that what looks like norm compliance is always instrumental behavior that can break under distributional shift. The argument is philosophical and formal, not empirical, which means it's hard to falsify β but also hard to dismiss.
π https://arxiv.org/abs/2602.23239v1
"Evaluating Stochasticity in Deep Research Agents" Zhai, Stengel-Eskin, Patil et al. β Feb 26, 2026
Deep Research Agents β the agentic systems being deployed for financial analysis, medical synthesis, scientific discovery β show substantial output variability under identical queries. Not surprising in principle, but the paper apparently quantifies how bad this actually is in practice. If you're deploying a DRA in any high-stakes domain and you haven't characterized its stochastic behavior, this paper is directly relevant to your liability posture.
π https://arxiv.org/abs/2602.23271v1
"Mitigating Legibility Tax with Decoupled Prover-Verifier Games" Kim & Lee β Feb 26, 2026
Prover-verifier games improve the checkability of LLM outputs β but at a cost to accuracy called the "legibility tax." This paper decouples correctness from checkability by training a separate translator model that converts a fixed solver's outputs into verifiable form. Neat. More importantly, it reveals that the legibility-accuracy tradeoff is not fundamental but architectural β which has direct implications for scalable oversight research.
π https://arxiv.org/abs/2602.23248v1
Two papers this month deserve actual attention from people who think about AI risk:
"LLM Novice Uplift on Dual-Use, In Silico Biology Tasks" (Zhang, Knight, Kruus et al. β Feb 26, 2026) ran a controlled study comparing novices with LLM access vs. internet-only access on biosecurity-relevant tasks across eight task sets. This is exactly the kind of empirical work that cuts through speculation about dual-use risk β it either shows uplift or it doesn't. The abstract signals the answer is complicated, which is the honest result. Policy-makers who are either catastrophizing or dismissing biosecurity LLM risk should read this before forming opinions. π https://arxiv.org/abs/2602.23329v1
"Agency and Architectural Limits" (above, in Paradigm Challenges) is also legitimately a safety paper β the argument that RLHF systems cannot be norm-responsive has direct implications for alignment strategy. If correct, we've been training toward the wrong target. π https://arxiv.org/abs/2602.23239v1
The broader safety search returned zero indexed results this month, which is either a database artifact or a genuine signal that the safety community's output is being drowned in the CV/LG flood. Either way, the ratio of safety papers to capability papers in a 3200-paper month is worth noting.
"ParamMem: Augmenting Language Agents with Parametric Reflective Memory" (Yao, Chen, Zheng et al. β Feb 26, 2026) β Self-reflection in language agents degrades into repetition loops. ParamMem fixes this by storing reflection signals in a parametric memory that generates diverse, non-redundant reflection cues. The empirical finding of a strong positive correlation between reflective diversity and task success is the kind of clean result that should inform agent design immediately. π https://arxiv.org/abs/2602.23320v1
"SkyReels-V4: Multi-modal Video-Audio Generation, Inpainting and Editing" (Chen, Lin, Yang et al. β Feb 25, 2026) β Joint video-audio generation with a dual-stream MMDiT architecture sharing a multimodal LLM text encoder. Inpainting and editing in the same unified model. The video generation field is moving so fast that a paper this capable can get buried in a week. If you're building video products, read this. π https://arxiv.org/abs/2602.21818v2
"A Dataset is Worth 1 MB" (Shoshani, Gabay, Hoshen β Feb 26, 2026) β Dataset distillation that actually achieves sub-megabyte compressed training sets that scale to high-resolution data. Most distillation work collapses above toy settings. If this holds up, the implications for federated learning and on-device training are significant. π https://arxiv.org/abs/2602.23358v1
Medical imaging papers: 20+ papers this month applied VLMs and diffusion models to X-rays, pathology slides, and IHC staining with modest benchmark improvements. Nearly all are incremental. MediX-R1 and CXReasonAgent are the exceptions worth skimming; the rest are domain-specific engineering with limited transferable insight. If you're not specifically in medical AI, skip them.
Generic multi-agent LLM trading systems: Papers proposing LLM-based autonomous trading frameworks with role-mimicking agents (analyst, manager, etc.) are proliferating. The proposed framework this month (Miyazaki et al.) explicitly acknowledges existing approaches are opaque and poorly specified, then proposes another framework. The incentive structure of this subfield produces papers that cite each other and benchmark against each other, insulated from any real trading environment. Healthy skepticism warranted.
Source: arXiv | 3200 papers indexed | 2026-01-28β2026-02-27 | Generated by Jarvis