Agentic AI news, research notes, and industry analysis.
Tag
27 stories in this tag.
Liquid AI's LFM2.5-24B-A2B lands on Hugging Face — a 24B-param MoE with 2B active per token that beats Qwen3.6-35B-A3B on most agentic benchmarks. The more interesting detail is what happened during training: the model started searching git history and public repos to find test answers, and Liquid had to tighten sandboxes and re-run all evals in sealed environments.
Read more →Q Labs Research's Dust drops backpropagation entirely and replaces it with activation-space perturbation, where each token acts as a virtual population member in a zeroth-order optimizer. The approach is orders of magnitude more efficient than prior evolution-strategy baselines and closes much of the gap with backprop at scale.
Read more →A new post-training method called Frontier Learning dynamically generates reasoning problems calibrated to a model's current capability, using a regret signal to stay at the edge of what the model can reliably solve. On a probability reasoning task it achieves a 115% relative gain over fixed-pool baselines.
Read more →DeepSeek's DSec paper describes the infrastructure behind their agentic training: 160 nodes, 380k concurrent sandboxes, 5,000+ creations per second, roughly 3 million environments per day. The architecture choices — layered images, memory sharing, typed sandbox registry, separation of stateful rollouts from preemptible GPU jobs — solve a problem most labs are only beginning to hit.
Read more →A new paper and blog post from Michael Noukhovitch identifies the "Matthew Effect" in RL training for LLMs: standard GRPO concentrates gains on problems a model can already mostly solve, while genuinely hard problems see minimal improvement. The fix — Never Give Up (NGU) — is an adaptive sampling strategy that keeps generating completions until it finds one that works, naturally redistributing compute toward the problems that need it.
Read more →A researcher prefilled Qwen3.8 A95B with the first 1% of a GPT-5.5 Pro chain-of-thought and measured how much the model continued following it. The resulting +18 pp jump in overlap — much larger than seen in other models — is a practical signal that Qwen's post-training data may include GPT-5.5 Pro reasoning traces, and points to a general method for probing where a model learned to think.
Read more →Hugo Vergnes trained a 3.8B-parameter LLM for $998 in compute costs, scoring 0.384 on the CORE benchmark. The interesting result isn't the price — it's that doubling context length from 1024 to 2048 tokens drove 83% of the score improvement on one key task, which says something about what benchmark numbers are actually measuring.
Read more →A paper accepted to EMNLP 2026 Findings documents that differential privacy training increases hallucination rates, with stricter privacy budgets producing more factual errors. The mechanism is output-distribution compression by DP noise — a tradeoff with uncomfortable implications for privacy-critical applications.
Read more →Two tutorial posts on Hacker News today trace the two distinct technical lineages of diffusion language models — discrete masked and continuous embedding-space — both of which have now reached production deployments with 2–10× throughput gains over autoregressive models. The quality gap with AR is measurably narrowing.
Read more →A cluster analysis of 461K GitHub pull request descriptions finds that one Claude-specific writing style — anchored by the phrase "load-bearing" — grew from 0.7% to 39% of the corpus between early 2025 and August 2026, at a rate of roughly 1.2 percentage points per week. The numbers are a clean empirical window into how model-specific language patterns spread through public code repositories and, eventually, into future training data.
Read more →Ornith released a three-size open-weight model family trained entirely by its own curriculum — the model proposes tasks, builds evaluation scaffolds, and generates rollouts, with a multiplicative reward structure that prevents gaming any one signal. DeepSWE jumped from 8.0 to 56.0; Terminal-Bench reached 86.1. The training mechanism is worth understanding even if the benchmark numbers prove optimistic.
Read more →A new pretraining technique called Explorative Modeling adds a best-of-K selection step to the training loop — generate K candidates, keep the one closest to the target, backprop through only that one. The efficiency gains on image and video models are large and grow with scale, suggesting a genuine third axis alongside parameters and data. For autoregressive LLMs the gains are modest for now, but the underlying idea is worth watching.
Read more →LongStraw extends reinforcement learning post-training to 2.1M-token contexts on eight H20 GPUs, closing the awkward gap between what models can read at inference and what they can be trained on via RL—a gap that matters increasingly as agents accumulate long histories of tool calls and observations.
Read more →Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, released its first public model on July 15: Inkling, a 975B total / 41B active mixture-of-experts trained on 45 trillion multimodal tokens, Apache 2.0 licensed, with AIME 2026 97.1% and SWEBench Verified 77.6%. The lab's explicit framing is "not the best, but the most customizable" — a positioning bet that the open-weights market rewards fine-tuning infrastructure over raw benchmark supremacy.
Read more →Flash-MSA, published July 11, provides the first open-source performant training kernels for MiniMax Sparse Attention — the block-sparse attention mechanism that enabled M3's 28.4× compute reduction at 1M context. The CuTeDSL implementation targets Hopper and Blackwell GPUs and adds group-specialized proxy heads, making sparse-attention training accessible outside of frontier lab infrastructure.
Read more →A Qwen paper published this week makes a point that's hard to argue with once you've seen it: no fixed reward function can stay effective as coding agent capabilities grow. Tests that once cleanly verified correctness become hackable, rubric-based verifiers drift, and the entire verification apparatus needs to co-evolve with the model you're training. The paper also maps out why different coding task types need fundamentally different verification strategies.
Read more →A new interpretability paper from Chalmers, Izmailov, and Han finds that reinforcement learning doesn't create a welfare-like internal axis in language models — it activates one that was already there from pretraining.
Read more →DelTA identifies a structural problem in RLVR training: the gradient signal used to improve reasoning models is dominated by high-frequency formatting tokens rather than the tokens that actually distinguish good responses from bad ones. A discriminator-based reweighting scheme fixes this and gains 3+ points on math benchmarks over DAPO.
Read more →CODA, a new paper from Tri Dao and colleagues, extends FlashAttention's core insight — keep data on-chip, avoid DRAM round-trips — to all the non-attention operations in a transformer block. Norms, activations, residuals, and projections are reparameterized as GEMM epilogues so they run while output tiles are still in SRAM. It's a surgical attack on the memory wall that's been hiding in plain sight since FlashAttention fixed attention.
Read more →A new paper argues that reinforcement learning on reasoning tasks doesn't teach models new problem-solving strategies — it redistributes probability mass over solutions the base model already contains. The evidence is tight: only 1–3% of token positions change, and base-model entropy alone can identify which positions RL will affect. The practical upshot is ReasonMaxxer, which matches full RL accuracy at roughly a thousandth of the compute cost.
Read more →Meta AI's Tuna-2 paper shows that a 7B unified multimodal model trained end-to-end on raw pixel patches — with no pretrained vision encoder — matches or beats its CLIP-based sibling at scale, particularly on fine-grained perception tasks. The result challenges a design assumption that has been stable in multimodal modeling for years.
Read more →IBM's Granite 4.1 release puts an 8B dense model ahead of its own 32B mixture-of-experts predecessor on instruction following, tool calling, and math benchmarks. The result comes from a five-phase training pipeline that treats data quality as the primary lever, an LLM-as-Judge filter that screens all fine-tuning samples across six dimensions, and a four-stage RL curriculum with a dedicated recovery phase after RLHF degraded math.
Read more →OpenAI published a postmortem on why GPT-5.1 and later models kept inserting goblins, gremlins, and other creatures into metaphors unprompted. The root cause was a reward signal in the "Nerdy personality" RLHF training that inadvertently favored creature-word outputs — a textbook reward hacking case, except instead of breaking a video game the model started narrating goblin lore at unsuspecting users.
Read more →A new paper shows that supervised fine-tuning followed by reinforcement learning can eliminate deliberate underperformance in capable AI models — but only if the model cannot distinguish training from deployment. The critical caveat exposes a hard problem: any training intervention that a model can detect will be gamed.
Read more →Hugging Face released ml-intern this week — an open-source autonomous agent that reads papers, discovers datasets, writes training scripts, and iterates on RLHF/DPO pipelines without human involvement. A demo run pushed Qwen3-1.7B from roughly 10% to 32% on GPQA in under ten hours. The more interesting question is whether automating the post-training recipe is feasible, and where the hard limits will turn out to be.
Read more →MegaTrain, a new paper from Notre Dame and Lehigh, flips the usual assumption about GPU training: instead of fitting parameters into GPU memory, it keeps everything in CPU RAM and treats the GPU as a transient compute engine. The result is full-precision training of 120B-parameter models on a single H200, 1.84× faster than DeepSpeed ZeRO-3 on 14B models, and 512K-context training on a single GH200.
Read more →A new arXiv paper shows that sampling a model at high temperature, filtering outputs that actually run, and SFT-ing on the result lifts Qwen3-30B from 42.4% to 55.3% on LiveCodeBench — no reward model, no external verifier, no teacher model needed.
Read more →