#Paper/Research

AIMost alignment research is not crisp and requires research taste to evaluate, according to Jan Leike. He explains that Anthropic chose to point the AAR at a scalable oversight problem because progress there could let AARs tackle fuzzier alignment problems where humans can only provide weak supervision.

AIJan Leike, owner of this account at Anthropic, praised the work of @jiaxinwen22, @liangqiu_1994, Joe Benton, and @janhkirchner. He linked to a blog post on Anthropic's research page about automated alignment researchers.

AIAnthropic's Claude developed scalable oversight methods on chat reward modeling datasets and evaluated them on math and code datasets. The best methods performed strongly on math but gave mixed results on code, suggesting the methods were overfit to the data and models used.

AICognition and Applied Compute RL-trained SWE-check, a specialized bug detection model for the Windsurf IDE. It matches frontier performance on in-distribution evals and is an order of magnitude faster with cheaper inference, though it trails frontier models on out-of-distribution evals (delta F1 0.29 versus 0.49 before training). A preview is available in Windsurf Next, with a mainstream release planned.

Why it matters: The post explains how production environment replication, reward linearization, and two-phase post-training trade bug-detection quality against latency for an IDE specialist model.

AIIntern Large Models introduced Kernel-Smith, a framework for generating high-performance GPU kernels and operators using an evolutionary agent and post-training recipe. The post says it outperforms Gemini-3.0-pro and Claude-4.6-opus on Kernel-Bench, and that optimized kernels have been merged into SGLang and LMDeploy. The accompanying figure compares best program score trajectories across evolution steps, with Kernel-Smith-235B-RL reaching the highest peak.

AIMistral has released Voxtral TTS, a text-to-speech model, alongside a blog post, a playground, a technical report, and model weights on Hugging Face. The post itself contains only links and no further details about the model's capabilities.

AIMistral's Voxtral TTS is its first speech model, presented as an open-weight text-to-speech model that reportedly delivers SOTA performance at significantly lower cost with very low latency. It combines autoregressive generation of semantic speech tokens with flow-matching for acoustic tokens, and a technical report on its training methodology is being released.

AIDataChef, an AI4AI framework, uses reinforcement learning to automatically generate optimal data recipes for adapting LLMs. Its DataChef-32B model, using an efficient proxy reward system, matches Gemini-3-Pro in recipe generation, with its recipes surpassing expert-curated ones on AIME'25 and ClimaQA benchmarks.

AIIntern Large Models introduced ARM-Thinker, an agentic multimodal reward model that autonomously invokes image crop and zoom-in, document retrieval, and instruction-following validators to ground its judgments in evidence.

AIARC Prize has released ARC-AGI-3, a set of hundreds of interactive, turn-based environments with thousands of game-style levels, with no instructions or stated goals. Humans score 100% while frontier AI scores 0.51%. ARC Prize 2026 offers over $2 million in prizes for open-source solutions to ARC-AGI-2 and ARC-AGI-3.

Why it matters: The benchmark's human versus frontier AI gap and its interactive design show how agent evaluation is shifting from instruction-following toward exploration and adaptation.

AITri Dao says nonlinear RNNs seem to do something genuinely different from attention and linear RNNs or SSMs. He reports they already perform well with the right parametrization, and adding just one nonlinear RNN layer substantially improves a transformer-Mamba/DeltaNet hybrid. The post quotes the M²RNN paper, which introduces non-linear RNNs with matrix-valued states for language modeling, with links to the paper, code, and models.

AIBAAI has released RoboBrain-Dex, a dexterous manipulation model for embodied intelligence trained on large-scale, diverse human egocentric data rather than massive robot teleoperation datasets. BAAI says this shifts robotic dexterous manipulation research from small data with weak generalization to big data with strong generalization. The code is open-sourced on GitHub.

AIBerkeley AI Research introduces SPEX, a signal-processing framework that identifies influential interactions in LLMs using far fewer ablations than exhaustive analysis. A hierarchy-based extension, ProxySPEX, matches SPEX performance with around 10x fewer ablations. The methods apply to feature, data, and model component attribution.

AIThe FlashAttention-4 paper is out, reporting that attention on Blackwell GPUs now runs at roughly matmul speed, reaching about 1600 TFLOPs. The forward pass is bottlenecked by exponential computation and the backward pass by shared memory bandwidth, and the redesign uses polynomial exponential emulation, a new online softmax that avoids 90% of softmax rescaling, and 2CTA MMA instructions that let two thread blocks share operands to cut shared memory traffic.

AITri Dao reposts a quoted post from @tanishqkumar07 introducing Speculative Speculative Decoding (SSD), an LLM inference algorithm claimed to be up to 2x faster than leading inference engines. The quoted post credits collaborators @tri_dao and @avnermay and links to a thread with details. Tri Dao's own text says the approach applies an asynchronous-machines principle seen in GPU kernels to speculative decoding.

AIAletheia, a math research agent, autonomously solved 6 of 10 FirstProof problems without modification, the best result in the inaugural challenge. The author says this is bigger than the IMO-gold achievement from last year, and the results were evaluated by experts with best-of-2 scoring.

AIResearchers trained a humanoid with 22-DoF dexterous hands mainly on over 20,000 hours of egocentric human video, with no robot in the loop, to perform tasks such as assembling model cars and folding shirts. They report a log-linear scaling law (R² = 0.998) between human video volume and action prediction loss, and state that this loss predicts real-robot success rate. The recipe, called EgoScale, pre-trains GR00T N1.5 on the video, adds only 4 hours of robot play data, and reports a 54% gain over training from scratch across five dexterous tasks.

AIGoogle researchers used the Aletheia agent, powered by Gemini 3 Deep Think, to attempt 10 FirstProof challenge problems without modification. The agent operated fully autonomously and solved 6 of the 10 problems, according to the post, with methodology and expert evaluations described in the linked arXiv paper.

Why it matters: The post gives the autonomous setup and expert-evaluated results for an AI agent on FirstProof math problems, useful for judging how far such systems go on research-level math.

AIMistral has released the Voxtral-Mini-4B-Realtime-2602 model weights on Hugging Face, alongside an arXiv paper and a live realtime audio playground in Mistral AI Studio. The post links these resources but provides no further benchmark figures or capability details.

AIMistral has published the technical report for Voxtral Realtime, a speech transcription model released under the Apache 2.0 license. The model reportedly achieves state-of-the-art transcription performance at sub-500ms latency. Mistral also launched a Realtime playground in Mistral Studio and made the model available in Hugging Face Transformers.

AIJakub Pachocki, owner of the OpenAI account, shared a PDF of solution attempts produced by OpenAI's model. The post links to the document but gives no further details on the problems, results, or methods.

AIMiniMax describes Forge, its internal reinforcement learning framework for training real-world agents, which was used during the development of MiniMax M2.5. The post explains a Windowed FIFO scheduler, prefix tree merging that the post says yields a 40x training speedup, and CISPO-based training across more than one hundred thousand agent scaffolds and environments.

Why it matters: The post details how the Forge framework balances throughput, stability, and agent flexibility, with concrete scheduling and prefix-merging methods for training agent RL at scale.

AIYi Tay introduces Aletheia, a math research agent powered by an advanced version of Gemini Deep Think. The post says it produced two publishable papers, one fully automatic and one human-AI collaboration, and solved multiple open Erdős problems. The attached image shows a Google DeepMind paper titled "Towards Autonomous Mathematics Research" with a generator, verifier, and reviser loop.

AIA new paper shows researchers using "vibe-proving" with Gemini Deep Think to tackle open problems and refute long-standing conjectures. The work also reports discovering obscure mathematical links across theoretical computer science, physics, and other fields.

AIQuoc Le announced a paper describing Aletheia, an agent built on Gemini Deep Think that goes beyond Olympiad problems to PhD-level mathematics. The post says Aletheia iteratively generates and verifies proofs, collaborates on human-AI research, autonomously generates a paper on eigenweights, and solves open Erdős conjectures.

AIAnthropic reports that resource configuration alone can move Terminal-Bench 2.0 scores by up to 6 percentage points, with infra error rates falling from 5.8% under strict enforcement to 0.5% when uncapped. Above about 3x the per-task specs, extra headroom starts letting agents solve tasks they previously could not, so limits can change what the eval measures.

Why it matters: The source shows how container resource limits shift agentic coding scores, which helps readers interpret small leaderboard gaps and set up evals more consistently.

AIQuoc Le announced a case study using Gemini to systematically evaluate 700 conjectures labeled open in the Erdős Problems database. The team addressed 13 problems, finding 5 novel autonomous solutions and identifying 8 existing solutions missed by previous literature.

AIThe Beijing Academy of Artificial Intelligence (BAAI) published its Emu3 multimodal large model research in Nature, which the post describes as the first large-model achievement led by a Chinese research institution in that journal. Emu3 learns from text, image, and video at scale using next-token prediction alone, reaching generation and perception performance comparable to task-specific methods. The authors frame this as a step toward scalable, unified multimodal intelligence systems.

AIAi2's Open Coding Agents family, with SERA as its first release, was built by Tim Dettmers and collaborators on 32 GPUs. The method generates synthetic bug trajectories with soft verification, comparing patches by line overlap instead of running tests. The post reports that a 32B model fine-tuned on about 7,000 trajectories for one private repository matched its GLM 4.5-Air teacher, and that the baseline costs $500 to run.

AIBerkeley AI Research proposes an information-based framework that evaluates and optimizes imaging systems using mutual information estimated directly from noisy measurements. The team reports that the metric predicts decoder performance across color photography, radio astronomy, lensless imaging, and microscopy, and that optimized designs match end-to-end methods while requiring less memory and compute.

AITsinghua AIR and BAAI's DrugCLIP screened 10,000 proteins against 500 million molecules, identifying over 2 million drug candidates. The post claims a 1-million-fold speedup, reaching 10 trillion protein-molecule pairs per day, and positions DrugCLIP as bridging AlphaFold structures to drug candidates. The work is published in Science, with a platform available at drugclip.com.

AIMiniMax released the technical report and pretrained weights for VTP-Large-f16d64, a visual tokenizer that jointly optimizes contrastive, self-supervised, and reconstruction losses. The model scores 78.2 zero-shot accuracy, 85.7 linear probing, and 0.36 rFID, and its generation performance scales with pretraining compute, parameters, and data. Checkpoint weights were listed as "released very soon" in the source.

AIApple has released SHARP, a model that generates a 3D Gaussian representation of a scene from a single photograph in less than a second on a standard GPU. The output renders in real time as high-resolution photorealistic views of nearby camera positions, with metric absolute scale, and the paper reports reductions of 25–34% in LPIPS and 21–43% in DISTS versus the best prior model.

AIOpenAI has published weights for a sparse model from Gao et al. 2025, used for qualitative results on bracket counting and variable binding, on Hugging Face under the openai/circuit-sparsity repository. The release includes a standalone Hugging Face implementation that loads the converted model and tokenizer with trust_remote_code and runs sample generation. The project is licensed under Apache License 2.0.

AIThinking Machines published a post on on-policy distillation, a training approach combining the error-correcting relevance of RL with the reward density of SFT. The quoted post reports that in math reasoning and an internal chat assistant, on-policy distillation can outperform other approaches at a fraction of the cost.

AIThinking Machines Lab describes on-policy distillation, which samples rollouts from a student model and has a teacher grade each token with reverse KL. The authors report that this matches Qwen3-style reasoning results at a fraction of RL's cost, with AIME'24 reaching 70% in about 150 steps from a 400k SFT checkpoint. The method also helps recover instruction-following behavior lost during fine-tuning on internal documents.

Why it matters: The post explains why on-policy distillation gives dense per-token feedback, letting a small model match RL results at much lower compute cost.

AIStanford and Cognition AI researchers introduced Kevin-32B, a 32B-parameter model trained with multi-turn reinforcement learning to write CUDA kernels. On KernelBench, it solves 89% of tasks at best@16 and achieves 65% average correctness over eight refinement steps, versus 53% for o4-mini and 51% for o3. Its best@16 speedup is 1.41x, and multi-turn training outperforms single-turn training as refinement steps increase.

AILiquid AI reports STAR, an evolutionary algorithm that synthesizes tailored neural network architectures from a numerical genome representation. The authors say it produced hundreds of designs that outperform Transformer and hybrid architectures in quality, with smaller caches and parameter counts, and can optimize for latency on target hardware. The full method is described in the arXiv technical report 2411.17800.

Why it matters: The post explains how evolutionary search over a new architecture design space produced designs beating Transformers and hybrids, giving a concrete method for quality versus latency and memory trade-offs.

AICognition reports that its agent Devin resolved 79 of 570 sampled SWE-bench issues, a 13.86% success rate, without being given the files to edit. The report says this exceeds the best previous unassisted baseline of 1.96% and the best assisted result of 4.80%. It also describes the adapted evaluation setup, a 45-minute runtime limit, and cases where Devin failed on multi-file edits.

Why it matters: The report explains how SWE-bench was adapted for end-to-end agent evaluation, with failure cases that clarify where the 13.86% result comes from and its limits.

AIA ByteDance research team won the Best Paper Award at the Association for Computational Linguistics (ACL) 2021 for VOLT, a simple and efficient machine translation solution. The team has open-sourced VOLT for developers to use.