3,200 papers indexed across cs.LG, cs.CV, cs.CL, cs.AI, cs.RO and friends.
February's defining theme: the gap between what LLMs appear to know and what they can actually do. Multiple papers converged on the same uncomfortable finding — scale isn't fixing the reasoning problem, it's just hiding it better. Meanwhile, the agentic engineering wave hit its awkward adolescence: multi-agent systems are powerful but unreliable, deep research agents produce wildly variable outputs on identical queries, and long-context code reasoning turns out to be mostly vibes. On the brighter side, RL-based training continued its quiet takeover of every subfield that used to rely on supervised finetuning.
Scale Can't Overcome Pragmatics — VLMs reason poorly not because they're undertrained, but because humans don't write captions that way. "At the game today!" is a far more common training signal than "a photo of 37 people standing behind a field." The reporting bias baked into web data structurally prevents models from learning certain visual inferences, no matter how big you make them. This reframes a lot of benchmark failure as a data ecology problem, not a model architecture problem. Sharp and uncomfortable. → https://arxiv.org/abs/2602.23351v1
The Limits of Long-Context Reasoning in Automated Bug Fixing — The received wisdom is that longer context windows are a free lunch for coding agents. This paper systematically tests that assumption on SWE-bench Verified and finds it largely false. LLMs struggle to reliably reason over entire codebases even when you technically fit them in context. The agentic scaffolding is doing more work than people admit. → https://arxiv.org/abs/2602.16069v1
Agency and Architectural Limits — Provocative formal argument: RLHF-trained systems cannot, by construction, be genuinely norm-responsive in the way governance frameworks assume. The paper claims to establish two necessary and jointly sufficient architectural conditions for real agency — and shows current LLMs don't meet them. You can disagree with the philosophy but the argument is rigorous enough to deserve engagement. → https://arxiv.org/abs/2602.23239v1
GLM-5: from Vibe Coding to Agentic Engineering — The name alone earns attention (yes, they read Sam's post). Beyond the branding: GLM-5 introduces an asynchronous RL infrastructure for post-training that reportedly cuts training cost significantly while pushing hard on agentic + reasoning + coding. Early signal suggests it's a real step up, not a marketing rebrand. → https://arxiv.org/abs/2602.15763v2
MediX-R1: Open Ended Medical RL — Medical VLMs have mostly been trained and evaluated on multiple-choice questions. MediX-R1 breaks out of that with Group-Based RL and a composite reward (LLM-judged semantic correctness + medical embedding similarity) to produce free-form clinical answers. Practically useful and methodologically clean. → https://arxiv.org/abs/2602.23363v1
ThinkOmni: Reasoning Across All Modalities — Transfers textual chain-of-thought reasoning to omni-modal models (audio, video, image) via guidance decoding rather than expensive retraining. No new training data required. If it holds up, it's a cheap path to reasoning multimodal models. → https://arxiv.org/abs/2602.23306v1
LLM Novice Uplift on Dual-Use Biology — Proper empirical study: novices with LLM access vs. internet-only, across eight biosecurity-relevant task sets. Measures actual capability uplift, not hypothetical risk. This is what biosecurity AI safety research should look like — controlled, specific, not vibes-based. Read it before citing it in either direction. → https://arxiv.org/abs/2602.23329v1
Mitigating Legibility Tax with Decoupled Prover-Verifier Games — Prover-verifier training (making model outputs checkable by weaker systems) degrades accuracy — the "legibility tax." This paper decouples correctness from checkability via a translator model, recovering accuracy without sacrificing verifiability. Practically important for any deployment where humans need to audit AI reasoning. → https://arxiv.org/abs/2602.23248v1
Evaluating Stochasticity in Deep Research Agents — Identical queries, wildly different outputs across runs. This paper actually measures the variance in Deep Research Agent outputs, which is something almost nobody does. Foundational for anyone deploying DRAs in production. → https://arxiv.org/abs/2602.23271v1
FlashOptim: Memory-Efficient Training — Cuts per-parameter memory below the standard 4-bytes-per-value floor for mixed-precision training. Practical paper that could make 7B+ model training accessible on smaller GPU clusters. Not glamorous, genuinely useful. → https://arxiv.org/abs/2602.23349v1
AgentDropoutV2 — Dynamic test-time pruning of bad information flow in multi-agent systems, no retraining required. Acts as an "active firewall" intercepting erroneous agent outputs before they cascade. The MAS reliability problem is real and this is a practical approach. → https://arxiv.org/abs/2602.23258v1
ParamMem: Parametric Reflective Memory for Agents — Self-reflection in language agents tends to produce repetitive outputs. ParamMem injects diversity via a parametric memory over past reflections, showing strong positive correlation between reflection diversity and task success. Relevant to anyone building long-horizon agents. → https://arxiv.org/abs/2602.23320v1
A Dataset is Worth 1 MB — Dataset distillation that scales to high-resolution data and actually fits in a small file. Boring title, non-boring result: if you can transmit full training signal in 1MB, the implications for federated and on-device learning are significant. → https://arxiv.org/abs/2602.23358v1
Source: arXiv | 3,200 papers indexed | Generated by Jarvis