LLMs

Yafu Li, Runzhe Zhan, Haoran Zhang, Shunkai Zhang, Yizhuo Li, Zhilin Wang, Jiacheng Chen, Futing Wang, Xuyang Hu, Yuchen Fan, Bangjie Xu, Yucheng Su, Xinmiao Han, Chenxi Li, Haodi Lei, Yufeng Zhao, Zejin Lin, Qianjia Cheng, Tong Zhu, Xiaoye Qu, Ganqu Cui, Peng Ye, Yun Luo, Zhouchen Lin, Yu Qiao, Bowen Zhou, Ning Ding, Yu Cheng

The paper describes a training recipe that turns a 30B-A3B post-trained reasoning model into an olympiad-level math and physics solver, using three stages: supervised fine-tuning ordered by reverse-perplexity curriculum, a two-stage RL pipeline (verifiable-reward RL followed by proof-quality RL with a generative reward model, self-refinement, and experience replay), and a test-time solve-verify-refine loop. The resulting model, SU-01, is trained on 338K sub-8K-token SFT trajectories plus 200 RL steps.

Dongrui Liu, Yu Li, Zhonghao Yang, Peng Wang, Guanxu Chen, Yuejin Xie, Qinghua Mao, Wanying Qu, Yanxu Zhu, Tianyi Zhou, Leitao Yuan, Zhijie Zheng, Qihao Lin, Yimin Wang, Haoyu Luo, Shuai Shao, Chen Qian, Qingyu Liu, Ling Tang, Ruiyang Qin, Qihan Ren, Junxiao Yang, Kun Wang, Zhiheng Xi, Linfeng Zhang, Ranjie Duan, Bo Zhang, Wenjie Wang, Wen Shen, Qiaosheng Zhang, Yan Teng, Chaochao Lu, Rui Mei, Man Li, Jialing Tao, Xi Lin, Tianhang Zheng, Yong Liu, Quanshi Zhang, Lei Zhu, Xingjun Ma, Junhua Liu, Hui Xue, Xiaoxiang Zuo, Xiangnan He, Chao Shen, Xianglong Liu, Minlie Huang, Jing Shao, Xia Hu

A framework for training small guardrail models (0.8B to 8B parameters) that audit an AI agent's full execution trajectory and label it safe or unsafe, plus diagnose the risk source, failure mode, and real-world harm. It combines a taxonomy-guided synthetic data engine, influence-function-based data selection to shrink training to about 1k samples, and two-stage SFT plus RL (using a per-dimension variant of GRPO called GDPO).

Xiangchen Cheng, Yunwei Jiang, Jianwen Sun, Zizhen Li, Chuanhao Li, Xiangcheng Cao, Yihao Liu, Fanrui Zhang, Li Jin, Kaipeng Zhang

The paper introduces AgenticSTS, an LLM agent for the deck-building game Slay the Spire 2 that replaces the usual growing chat transcript with a bounded memory contract: every decision prompt is rebuilt fresh from five typed slots (protocol, state schema, game rules, episodic summaries, and triggered strategic skills) via typed retrieval, so no raw cross-decision history is appended. It also ships a testbed of 298 tagged trajectories, frozen memory snapshots, and analysis scripts so individual memory layers can be ablated in isolation.

Wei Zhou, Xuanhe Zhou, Shaokun Han, Hongming Xu, Guoliang Li, Zhiyu Li, Feiyu Xiong, Fan Wu

A benchmark study that decomposes 12 LLM agent memory systems (MemGPT/Letta, Mem0, Zep, A-MEM, MemOS, MemoryOS, Cognee, etc.) into four modules (representation/storage, extraction, retrieval/routing, and maintenance) and evaluates them from a data management angle. It runs end-to-end and ablation experiments across five workloads spanning 11 datasets, measuring not just task accuracy but retrieval fidelity, update robustness, long-horizon stability, and operational cost (index build time, query latency).

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin

This paper introduces the Transformer, a sequence-to-sequence architecture for machine translation that replaces recurrent and convolutional layers entirely with stacked self-attention and position-wise feed-forward layers. It uses scaled dot-product attention with multiple attention heads, plus sinusoidal positional encodings to inject word order since there is no recurrence.

Jacob Devlin, Ming-Wei Chang, Kenton Lee, Kristina Toutanova

BERT is a language representation model that pre-trains a deep bidirectional Transformer encoder using two self-supervised tasks: masked language modeling (predicting randomly hidden tokens from both left and right context) and next sentence prediction. The pre-trained model is then fine-tuned end-to-end for specific tasks by adding a single output layer, with no task-specific architecture changes.

Hao Zhang, Yiming Hu, Yong Wang, Mingqiao Mo, Xin Xiao, Xiangxiang Chu

BlockPilot is a lightweight predictor that picks a per-sample block size for diffusion-based speculative decoding, replacing the fixed block size inherited from training. It uses the target model's last-token predictive distribution after prefilling as input to a two-layer MLP classifier that selects the block size from a small local range around the training value.

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, Denny Zhou

The paper introduces chain-of-thought prompting: instead of giving a large language model few-shot examples that map questions directly to answers, each exemplar includes the intermediate reasoning steps that lead to the answer. The model then generates its own step-by-step reasoning before producing a final answer, with no finetuning involved.

Liliana Hotsko, Yinxi Li, Yuntian Deng, Pengyu Nie

Code2LoRA is a hypernetwork that reads a code repository (compressed into embeddings by a frozen Qwen3-Embedding model) and generates a repository-specific LoRA adapter for a frozen code LLM in a single forward pass, so repo knowledge lives in parameters instead of the prompt. It has two variants: Static, which maps one repository snapshot to an adapter, and Evo, which uses a GRU to update a hidden state per commit diff so the adapter tracks an evolving codebase.

Xuying Ning, Katherine Tieu, Dongqi Fu, Tianxin Wei, Zihao Li, Yuanchen Bei, Jiaru Zou, Mengting Ai, Zhining Liu, Ting-Wei Li, Lingjie Chen, Yanjun Zhao, Ke Yang, Bingxuan Li, Cheng Qian, Gaotang Li, Xiao Lin, Zhichen Zeng, Ruizhong Qiu, Sirui Chen, Yifan Sun, Xiyuan Yang, Ruida Wang, Rui Pan, Chenyuan Yang, Dylan Zhang, Liri Fang, Zikun Cui, Yang Cao, Pan Chen, Dorothy Sun, Ren Chen, Mahesh Srinivasan, Nipun Mathur, Yinglong Xia, Hong Li, Hong Yan, Pan Lu, Lingming Zhang, Tong Zhang, Hanghang Tong, Jingrui He

A survey that frames code as the 'harness' for LLM agents: the executable, inspectable, and stateful layer through which agents reason, act, model their environment, and coordinate. It organizes existing work into three layers (harness interface, harness mechanisms like planning/memory/tool-use/feedback control, and multi-agent scaling) and catalogs methods across coding assistants, GUI/OS automation, embodied agents, and scientific discovery.

Tianyi Zhou, Dongrui Liu, Leitao Yuan, Jing Shao, Xia Hu

COLLEAGUE.SKILL is a pipeline that takes heterogeneous traces from a person or role (chat logs, docs, review comments, interviews) and distills them into a versioned skill package compatible with the Agent Skills standard (a folder centered on SKILL.md). Each package splits into a capability track (procedures, review criteria, decision heuristics) and a bounded behavior track (communication style, interaction rules, correction log), and supports inspection, natural-language correction, rollback, and installation across agent hosts.

Hongcan Guo, Qinyu Zhao, Yian Zhao, Shen Nie, Rui Zhu, Qiushan Guo, Feng Wang, Tao Yang, Hengshuang Zhao, Guoqiang Wei, Yan Zeng

CoLa DLM is a text generation model that replaces left-to-right autoregression with a three-part pipeline: a Text VAE maps text to continuous latent variables, a block-causal Diffusion Transformer (DiT) learns a semantic prior over those latents via Flow Matching, and a conditional decoder turns generated latents back into text. The key design choice is that diffusion transports a latent prior (global semantics) rather than recovering token-level observations, separating semantic planning from token realization.

Kaiyi Zhang, Wei Wu, Yankai Lin

The paper reframes RLVR (reinforcement learning from verifiable rewards) policy updates as an implicit linear discriminator over token-gradient vectors, arguing that standard sequence-level methods like DAPO let shared high-frequency tokens (formatting, entities) dominate the update direction. It proposes DelTA, which computes per-token coefficients based on how well each token's gradient distinguishes positive-advantage from negative-advantage responses, then uses those coefficients to reweight the DAPO surrogate loss.

Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, Chelsea Finn

This paper introduces Direct Preference Optimization (DPO), a method for aligning language models to human preferences using a single binary cross-entropy classification loss instead of the standard RLHF pipeline. It shows that the KL-constrained reward maximization objective used in RLHF has a closed-form optimal policy, which lets you reparameterize the reward as a function of the policy itself and skip both the separate reward model and the reinforcement learning loop.

Guochao Jiang, Jingyi Song, Guofeng Quan, Chuzhan Hao, Guohua Liu, Yuewei Zhang

DVAO is a method for combining multiple reward signals in GRPO-style RL training of LLMs. Instead of using fixed weights to combine rewards or advantages, it scales each objective's contribution by that objective's empirical reward standard deviation within a rollout group, up-weighting objectives that show more variation (a stronger learning signal) and down-weighting noisy or saturated ones.

Tomas Mikolov, Kai Chen, Greg Corrado, Jeffrey Dean

This is the original word2vec paper. It introduces two log-linear model architectures, Continuous Bag-of-Words (CBOW) and Skip-gram, that learn dense word vectors by removing the expensive non-linear hidden layer used in prior neural language models and training on billions of words.

Jincheng Zhong, Weizhi Wang, Che Jiang, Kai Tian, Zhenzhao Yuan, Junlin Yang, Dianqiao Lei, Kaiyan Zhang

EnterpriseClawBench is a benchmark built by an automated pipeline that converts real internal workplace agent sessions into 852 reproducible tasks, each with recovered input fixtures, rewritten single-turn prompts, role/skill taxonomy labels, hard delivery rules, and semantic rubrics. It evaluates harness-model combinations (Claude Code, Codex, DeepAgents, Hermes, OpenClaw paired with various models) on producing usable business artifacts, scoring both objective delivery rules and LLM-judged quality across five dimensions.

Jundong Xu, Qingchuan Li, Jiaying Wu, Yihuai Lan, Shuyue Stella Li, Huichi Zhou, Bowen Jiang, Lei Wang, Jun Wang, Anh Tuan Luu, Caiming Xiong, Hae Won Park, Bryan Hooi, Zhiyuan Hu

EvoArena is a benchmark suite that turns static agent benchmarks (terminal workflows, software repos, user preferences) into version chains where the same environment evolves across releases, and the agent must adapt to changes while preserving still-valid prior behavior. The paper also proposes EvoMem, a git-like memory layer that logs non-additive memory updates as append-only patches (storing before/after state, rationale, and triggering evidence) and retrieves relevant patches alongside the latest memory at inference.

Zhilin Wang, Han Song, Runzhe Zhan, Jusen Du, Jiacheng Chen, Tianle Li, Qingyu Yin, Yulun Wu, Zhennan Shen, Tong Zhu, Yanshu Li, Guanjie Chen, Derek F. Wong, Yafu Li, Yu Cheng, Yang Yang

EvoPolicyGym is a benchmark that measures how well a coding agent (a language model plus its tool-using harness) can iteratively rewrite an executable decision policy for a reinforcement learning environment, using feedback from sandboxed rollouts under a fixed budget of 128 episodes. Agents edit a Python policy that exposes reset and act(obs), submit it for scored train rollouts, and get judged on held-out return of the checkpoint the server picks via hidden validation.

Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, Christopher Ré

FlashAttention is an exact attention algorithm that restructures the computation to avoid writing the large N x N attention matrix to GPU high bandwidth memory (HBM). It uses tiling to compute softmax block by block in fast on-chip SRAM and recomputes the attention matrix during the backward pass instead of storing it, all fused into a single CUDA kernel.

Alireza Salemi, Chang Zeng, Atharva Nijasure, Jui-Hui Chung, Razieh Rahimi, Fernando Diaz, Hamed Zamani

GrepSeek trains a compact LLM (Qwen3.5-9B) to answer questions by issuing Unix shell commands (rg, grep, head, awk, etc.) directly against a raw text corpus instead of querying a precomputed retrieval index. It uses a two-stage pipeline: supervised fine-tuning on a synthetic cold-start dataset built by an answer-aware Tutor and answer-blind Planner, followed by GRPO reinforcement learning, plus a sharded-parallel execution engine that runs shell pipelines across corpus shards while preserving byte-exact output.

Pengcheng Jiang, Zhiyi Shi, Kelly Hong, Xueqiang Xu, Jiashuo Sun, Jimeng Sun, Hammad Bashir, Jiawei Han

Harness-1 is a 20B search agent (built on gpt-oss-20b) trained with reinforcement learning inside a stateful harness that moves search bookkeeping out of the model and into the environment. The environment maintains a candidate pool, an importance-tagged curated set, an evidence graph, verification records, and deduplicated/compressed observations, while the policy only decides what to search, which documents to keep, what to verify, and when to stop.

Guan Wang, Changling Liu, Chenyu Wang, Cai Zhou, Yuhao Sun, Yifei Wu, Shuai Zhen, Luca Scimeca, Yasin Abbasi Yadkori

HRM-Text replaces the standard decoder-only Transformer with a Hierarchical Recurrent Model that splits computation into a slow strategic (H) module and a fast execution (L) module, stabilized for language via two tricks the authors call MagicNorm (a hybrid Pre/Post normalization scheme) and a warmup schedule for truncated backpropagation through time. Instead of pretraining on raw web text, it trains a 1B-parameter model from scratch only on instruction-response pairs, computing loss just on the response (a task-completion objective) with a PrefixLM attention mask that lets instruction tokens attend bidirectionally.

Nahyun Lee, Dongkeun Yoon, Guijin Son, Geewook Kim, Dayoon Ko, Jeonghun Park, Haneul Yoo, Jaewon Cho, Junghun Park, Changyoon Lee, Kyochul Jang, Jaeyeon Kim, Eunsu Kim, Woojin Cho, Seungone Kim

K-BrowseComp is a web-browsing agent benchmark of 400 problems grounded in Korean web contexts: a 300-question human-verified subset built by native Korean annotators, plus a 100-question synthetic split generated by a browsing agent (Claude Code) that constructs questions backwards from seed pages and filters them against a taxonomy of failure modes. Each question has a single stable answer requiring multi-hop or parallel-constraint retrieval across multiple Korean websites, and models are evaluated pass@1 under a fixed 10-call search budget.

Lorenz K. Muller, Philippe Bich, Chiara Boretti, Hyun-Min Chang, Jiawei Zhuang, Lukas Cavigelli

KVarN is a calibration-free KV cache quantizer that combines a Hadamard rotation in the channel dimension with an iterative Sinkhorn-style variance normalization across both the token and channel axes of the K and V matrices before round-to-nearest quantization. It targets the specific error accumulation that happens when the cache is quantized on-the-fly during long autoregressive decoding, rather than the parallel prefill setting most prior methods were tuned for.

Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, Dario Amodei

This is the GPT-3 paper: the authors train an autoregressive transformer language model with 175 billion parameters and evaluate it on over two dozen NLP tasks using in-context learning, where the model gets task demonstrations in its prompt at inference time with no gradient updates. They test three settings (zero-shot, one-shot, and few-shot with 10 to 100 examples in the context window) across eight model sizes to measure how performance scales.

Yifu Luo, Zeyu Chen, Haoyu Wang, Xinhao Hu, Yuxuan Zhang, Zhizhou Sha, Shiwei Liu

The paper introduces d-OPSD, an on-policy self-distillation method for diffusion language models (dLLMs) where a single model acts as both student and teacher. The teacher gets privileged information by revealing part of the model's own already-generated answer as a suffix condition, and training minimizes step-level KL divergence between student and teacher across the iterative denoising steps instead of the token-level divergence used for autoregressive models.

Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, Guillaume Lample

LLaMA is a family of transformer language models (7B to 65B parameters) trained only on publicly available datasets, using architectural changes like pre-normalization with RMSNorm, SwiGLU activations, and rotary positional embeddings. The models are trained on far more tokens than Chinchilla scaling laws recommend (up to 1.4T tokens), trading extra training compute for cheaper inference.

Jian Yang, Shawn Guo, Wei Zhang, Tianyu Zheng, Yaxin Du, Haau-Sing Li, Jiajun Wu, Yue Song, Yan Xing, Qingsong Cai, Zelong Huang, Chuan Hao, Ran Tao, Xianglong Liu, Wayne Xin Zhao, Mingjie Tang, Weifeng Lv, Ming Zhou, Bryan Dai

The paper studies how many times to loop a Parallel Loop Transformer (PLT), a looped-Transformer variant that reuses shared blocks with cross-loop position offsets (CLP) and shared-KV gated sliding-window attention to keep latency and KV-cache memory flat regardless of loop count. It trains LoopCoder-v2, a 7B coder, from scratch on 18T tokens at loop counts of 1, 2, 3, and 4, and uses per-loop diagnostics (hidden-state dynamics, attention routing, output-distribution shift) to explain why performance peaks at two loops.

Hongyuan Adam Lu, Z. L. Victor Wei, Qun Zhang, Jinrui Zeng, Bowen Cao, Lingwei Meng, Mocheng Li, Zezhong Wang, Haonan Yin, Naifu Xue, Minyu Chen, Cenyuan Zhang, Zefan Zhang, Hao Wei, Jiawei Zhou, Haoran Xu, Hao Yang, Ronglai Zuo, Tongda Xu, Yonghao Li, Jian Chen, Hebin Wang, Zeyu Gao, Yang Li, Wei Zhao, Qimin Zhong, Siqi Liu, Yumeng Zhang, Leyan Cui, Zhangyu Wang, Wai Lam

The paper applies looped transformers (a shared transformer block iterated repeatedly across depth) to world models, calling the result LoopWM. It adds a spectrally constrained state-retention matrix to keep latent updates bounded, Poisson-sampled variable loop depth during training, an early-exit gate for adaptive inference, and a deferred-decoding mode that skips intermediate observation reconstruction during multi-step rollouts.

Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen

LoRA (Low-Rank Adaptation) freezes a pre-trained model's weights and injects trainable rank-decomposition matrices (a pair of small matrices B and A whose product approximates the weight update) into each Transformer layer, so only those small matrices are trained. In their Transformer experiments the authors apply it mostly to the query and value attention projection matrices.

Jiacheng Chen, Xinyu Zhang, Shunkai Zhang, Yanmohan Wang, Lin Li, Tiancheng Qin, Qin Wang, Zhengmao Zhu, Tianle Li, Jingyang Li, Zehan Li, Binyang Jiang, Jin Zhu, Han Ding, Fei Yu, Chenyu Du, Zijian Song, Jiayuan Song, Zhi Zhang, Yunan Huang, Weiyu Cheng, Pengyu Zhao, Yu Cheng

MaxProof is a test-time scaling framework for competition math proofs that treats one merged model as generator, verifier, refiner, and ranker, then runs an evolution-style population search (dual PATCH/REWRITE refinement, conservative min-score fitness, pairwise tournament final selection). The underlying MiniMax-M3 model is trained through three specialist stages (proof generation via CISPO RL under a four-layer generative verifier, error-finding verification, and critique-conditioned repair) before being merged into a single release model.

Ye Jin, Yangyang Xu, Jun Zhu, Yibo Yang

MemSlides is an agent framework for generating and editing presentation slides that splits memory into three parts: long-term user profile memory (persistent style/layout preferences per user and intent), tool memory (reusable execution traces for editing tools), and session-scoped working memory (temporary constraints across revision turns). It pairs this with a Plan-Act-Guard revision pipeline that scopes edits to the smallest affected slide region rather than regenerating the whole deck on each feedback turn.

Mind Lab, :, Song Cao, Vic Cao, Andrew Chen, Kaijie Chen, Cleon Cheng, Steven Chiang, Kaixuan Fan, Hera Feng, Huan Feng, Arthur Fu, Jun Gao, Hongquan Gu, Aaron Guan, Nolan Ho, Mutian Hong, Hailee Hou, Peixuan Hua, Charles Huang, Miles Jiang, Nora Jiang, Yuyi Jiang, Qiuyu Jin, Fancy Kong, Andrew Lei, Kyrie Lei, Alexy Li, Lucian Li, Ray Li, Theo Li, Zhihui Li, Jiayi Lin, Kairus Liu, Kieran Liu, Logan Liu, Xiang Liu, Irvine Lu, Maeve Luo, Runze Lv, Pony Ma, Verity Niu, Anson Qiu, Vincent Wang, Rio Yang, Maxwell Yao, Carrie Ye, Regis Ye, Wenlin Ye, Josh Ying, Danney Zeng, Yuhan Zhan, Anya Zhang, Di Zhang, Ruijia Zhang, Sueky Zhang, Ya Zhang, Wei Zhao, Ada Zhou, Changhai Zhou, Yuhua Zhou, Xinyue Zhu, Murphy Zhuang

MinT is a managed infrastructure system for LoRA post-training and serving that keeps expensive base models resident on GPUs and moves only exported LoRA adapter revisions (not merged full checkpoints) through training, export, serving, and rollback. It provides a Tinker-compatible service API on top of vLLM samplers and PEFT/Megatron trainers, with a policy-record abstraction that tracks each adapter's base version, rank, checkpoint, and cache tier.

Disen Lan, Jianbin Zheng, Yuxi Ren, Xin Xia, Xuanda Wang, Xuefeng Xiao, Xipeng Qiu, Yu Cheng

FlashMorph is a method for choosing which layers of a pretrained Transformer to keep as full attention when converting the rest to linear attention (a hybrid model). Instead of fixed placement rules or scoring layers one at a time, it attaches a frozen linear-attention branch to each layer, adds a learnable scalar gate per layer, and jointly optimizes all gates on synthetic passkey-retrieval data with a regularizer that pushes layers toward linear attention, then keeps the top-K layers by gate value.

Yuru Wang, Lejun Cheng, Yuxin Zuo, Sihang Zeng, Bingxiang He, Che Jiang, Junlin Yang, Yuchong Wang, Kaikai Zhao, Weifeng Huang, Kai Tian, Zhenzhao Yuan, Jincheng Zhong, Weizhi Wang, Ning Ding, Bowen Zhou, Kaiyan Zhang

NatureBench is a benchmark of 90 tasks distilled from Nature-family science papers (2022-2025), built by an automated pipeline (NatureGym) that turns each paper into a containerized package with a task brief, data, a hidden test set, and an automated evaluator, while removing the original method so agents must find their own solution. It scores coding agents against each paper's published state of the art using a SOTA-normalized relative gap metric, with a post-hoc judge that flags shortcuts like output fabrication and feedback gaming.

Mind Lab, :, Vin Bo, Song Cao, Vic Cao, Andrew Chen, Kaijie Chen, Cleon Cheng, Steven Chiang, Kaixuan Fan, Hera Feng, Huan Feng, Arthur Fu, Jun Gao, Hongquan Gu, Aaron Guan, Nolan Ho, Mutian Hong, Hailee Hou, Peixuan Hua, Charles Huang, Miles Jiang, Nora Jiang, Yuyi Jiang, Qiuyu Jin, Fancy Kong, Andrew Lei, Kyrie Lei, Alexy Li, Lucian Li, Ray Li, Theo Li, Wenhao Li, Zhihui Li, Allen Lin, Jiayi Lin, Kairus Liu, Kieran Liu, Logan Liu, Xiang Liu, Irvine Lu, Maeve Luo, Runze Lv, Pony Ma, Verity Niu, Anson Qiu, Vincent Wang, Rio Yang, Maxwell Yao, Carrie Ye, Regis Ye, Wenlin Ye, Josh Ying, Danney Zeng, Yuhan Zhan, Anya Zhang, Di Zhang, Ruijia Zhang, Shiyang Zhang, Sueky Zhang, Ya Zhang, Wei Zhao, Ada Zhou, Adrian Zhou, Yuhua Zhou, Xinyue Zhu, Murphy Zhuang

A position paper from Mind Lab arguing that LoRA adapters should be treated as persistent per-user state on top of shared foundation models, organized around three axes: scaling up the base model (trillion-parameter MoE LoRA RL), scaling down the adapter (rank studies, a minor-subspace initialization called OLoRA-tail), and scaling out to many adapters (memory capacity laws, user simulation, majority voting). It bundles a range of experiments plus a serving/lifecycle system (MinT) that manages adapter revisions through storage, CPU cache, and GPU batch tiers.

Yuwen Du, Rui Ye, Shuo Tang, Keduan Huang, Xinyu Zhu, Yuzhu Cai, Siheng Chen

The paper trains a 30B search agent (OpenSeeker-v2, based on Qwen3-30B-A3B-Thinking) using only supervised fine-tuning on synthetic ReAct trajectories, skipping the usual continual pre-training and reinforcement learning stages. The main contribution is three tweaks to the data synthesis pipeline: expanding the knowledge graph size used to generate questions, enlarging the tool set, and filtering out any trajectory solvable in too few tool-call steps to enforce a difficulty floor.

Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, Jeff Dean

The paper introduces the Sparsely-Gated Mixture-of-Experts (MoE) layer, a neural network component containing up to thousands of feed-forward expert sub-networks where a trainable gating network selects only a few experts to run per input token. It uses noisy top-k gating plus auxiliary load-balancing losses, and applies the layer between stacked LSTM layers for language modeling and machine translation.

Caixin Kang, Tianyu Yan, Sitong Gong, Mingfang Zhang, Liangyang Ouyang, Ruicong Liu, Bo Zheng, Huchuan Lu, Kaipeng Zhang, Yoichi Sato, Yifei Huang

The paper introduces Grounded Personality Reasoning (GPR), a task that forces multimodal LLMs to justify each Big Five (OCEAN) personality rating with cited, timestamped behavioral evidence, and releases MM-OCEAN, a benchmark of 1,104 videos and 5,320 cue-grounding multiple-choice questions built by a multi-agent annotation pipeline with human verification. It benchmarks 27 MLLMs across three tiers (rating, open-ended reasoning, structured cue grounding) plus four failure-mode metrics that separate 'right answer' from 'right reasons'.

Wentao Zhang, Liliana Hotsko, Woojeong Kim, Pengyu Nie, Stuart Shieber, Yuntian Deng

The paper introduces Program-as-Weights (PAW), a system that compiles a natural-language description of a fuzzy task (things like log alerting, JSON repair, or intent classification that resist clean rule-based code) into a small neural artifact. A 4B compiler reads the spec plus an auto-generated pseudo-program of paraphrase-and-examples, then emits a per-function LoRA adapter that gets hot-attached to a frozen 0.6B interpreter model, which runs the task locally.

Yuxin Zuo, Zikai Xiao, Li Sheng, Fei Huang, Jianhong Tu, Yuxuan Liu, Tianyi Tang, Xiaomeng Hu, Yang Su, Qingfeng Lan, Yantao Liu, Qin Zhu, Yinger Zhang, Bowen Yu, Haiquan Zhao, Haiyang Xu, Jianxin Yang, Jiayang Cheng, Junyang Wang, Lianghao Deng, Mingfeng Xue, Tianyi Bai, Yang Fan, Yubo Ma, Yucheng Li, Zeyu Cui, Zhihai Wang, Zhihui Xie, Zhuorui Ye, An Yang, Dayiheng Liu, Jingren Zhou, Ning Ding

Qwen-AgentWorld is a language model trained to act as a world model: given interaction history and an agent's action, it predicts the next environment observation (tool response, terminal output, UI accessibility tree, etc.) across seven domains via chain-of-thought reasoning. It is built through a three-stage pipeline (continual pre-training, supervised fine-tuning, and RL with combined rubric-and-rule rewards) and evaluated on AgentWorldBench, a benchmark of real trajectories from frontier agents.

Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela

This paper introduces Retrieval-Augmented Generation (RAG), a fine-tuning recipe that pairs a pre-trained seq2seq generator (BART-large) with a neural retriever (Dense Passage Retriever) that pulls passages from a dense vector index of Wikipedia. It presents two variants: RAG-Sequence, which conditions the whole output on one retrieved document, and RAG-Token, which can draw a different document per generated token, with both retriever and generator trained end-to-end using only input/output pairs.

Lei Bai, Zongsheng Cao, Yang Chen, Zhiyao Cui, Shangheng Du, Yue Fan, Shiyang Feng, Zijie Guo, Haonan He, Liang He, Xiaohan He, Shuyue Hu, Yusong Hu, Songtao Huang, Yichen Jiang, Hao Li, Xin Li, Dahua Lin, Weihao Lin, Fenghua Ling, Dongrui Liu, Zhuo Liu, Runmin Ma, Chunjiang Mu, Haoyang Peng, Tianshuo Peng, Jinxin Shi, Luohe Shi, Boyuan Sun, Zelin Tan, Shengji Tang, Qianyi Wang, Yiming Wu, Yi Xie, Xiangchao Yan, Jingqi Ye, Peng Ye, Fangchen Yu, Jiakang Yuan, Bihao Zhan, Bo Zhang, Chen Zhang, Shufei Zhang, Shuaiyu Zhang, Wenlong Zhang, Yiqun Zhang, Junpeng Zhao, Zhijie Zhong, Bowen Zhou, Yuhao Zhou

A recipe for training a 35B mixture-of-experts agent (Agents-A1) to handle long-horizon agentic tasks by scaling trajectory length rather than parameters. The method has three stages: full-domain supervised fine-tuning on ~100K long trajectories (average 45K tokens), training domain-specific teacher models, then consolidating them into one student via domain-routed on-policy distillation with salient vocabulary alignment (SVA), which aligns student and teacher on the teacher's top-k token support instead of just the sampled token.

Zhengxi Lu, Zhiyuan Yao, Zhuowen Han, Zi-Han Wang, Jinyang Wu, Qi Gu, Xunliang Cai, Weiming Lu, Jun Xiao, Yueting Zhuang, Yongliang Shen

SDAR trains multi-turn LLM agents by combining GRPO reinforcement learning with an auxiliary on-policy self-distillation loss, where a teacher branch is the same policy given privileged training-only context (retrieved skills). The distillation signal is filtered through a per-token sigmoid gate that strengthens supervision on tokens the teacher endorses (positive teacher-student log-probability gap) and softly attenuates tokens where the teacher disagrees.

Ilya Sutskever, Oriol Vinyals, Quoc V. Le

This paper introduces the sequence-to-sequence (seq2seq) encoder-decoder architecture: one multilayer LSTM reads an input sentence into a fixed-length vector, and a second LSTM decodes that vector into the output sentence one word at a time. It applies this to English-to-French machine translation and adds a key trick of reversing the word order of the source sentence before feeding it in.

Yifan Yang, Ziyang Gong, Weiquan Huang, Qihao Yang, Ziwei Zhou, Zisu Huang, Yan Li, Xuemei Gao, Qi Dai, Bei Liu, Kai Qiu, Yuqing Yang, Dongdong Chen, Xue Yang, Chong Luo

SkillOpt is a text-space optimizer that treats an agent's natural-language skill document as a trainable object, borrowing deep-learning controls like batches, a textual learning rate (bounded number of edits per step), a held-out validation gate, and epoch-wise slow/meta updates. A frontier optimizer model reads execution trajectories, proposes structured add/delete/replace edits to the skill, and only accepts edits that improve validation performance, exporting a compact best_skill.md (300 to 2,000 tokens) while the target model and harness stay frozen.

Hongyi Liu, Haoyan Yang, Tao Jiang, Bo Tang, Feiyu Xiong, Yuyu Luo, Zhiyu Li

SkillsVote is a lifecycle framework for managing libraries of Agent Skills (structured folders of instructions, scripts, and applicability conditions used by LLM coding agents). It does three things: performs agentic search over skill folders to pick relevant skills before a task, decomposes execution traces into skill-linked subtasks after a task, and gates library updates so only successful, reusable discoveries edit or create skills.

Jiajie Jin, Yuyang Hu, Kai Qiu, Qi Dai, Chong Luo, Guanting Dong, Xiaoxi Li, Tong Zhao, Xiaolong Ma, Gongrui Zhang, Zhirong Wu, Bei Liu, Zhengyuan Yang, Linjie Li, Lijuan Wang, Hongjin Qian, Yutao Zhu, Zhicheng Dou

Arbor is a framework for autonomous research (which the authors call Autonomous Optimization) where an LLM agent iteratively improves a code artifact against a fixed objective without step-level human supervision. Its core mechanism is Hypothesis Tree Refinement (HTR): a persistent tree whose nodes bind a hypothesis, a git-worktree artifact version, evaluation evidence, and a distilled insight, managed by a long-lived coordinator that dispatches short-lived executors and only promotes changes that pass a held-out test evaluator.

Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katie Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Osindero, Karen Simonyan, Erich Elsen, Jack W. Rae, Oriol Vinyals, Laurent Sifre

The paper studies how to split a fixed training compute budget between model size and number of training tokens, fitting three empirical estimators (minimum-over-training-curves, IsoFLOP profiles, and a parametric loss function) across over 400 models trained from 70M to 16B parameters. It finds model size and training tokens should scale in roughly equal proportion (doubling one means doubling the other), then validates this by training Chinchilla, a 70B model on 1.4T tokens, at the same compute as the 280B Gopher.

Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, Ryan Lowe

This is the InstructGPT paper, which fine-tunes GPT-3 to follow instructions using reinforcement learning from human feedback (RLHF). The pipeline has three stages: supervised fine-tuning on labeler-written demonstrations, training a reward model on human rankings of model outputs, and then optimizing the policy against that reward model using PPO.

Hanyu Guo, Jiedong Yang, Chao Chen, Longfei Xu, Kaikui Liu, Xiangxiang Chu

TransitLM is a dataset of over 13 million transit route planning records (from Amap logs across four Chinese cities, 120,845 stations) plus a three-task benchmark, released as a continual pre-training corpus and supervised fine-tuning data. The authors validate it by continually pre-training then fine-tuning Qwen3 models (0.6B to 4B) that generate complete transit routes end-to-end from origin-destination GPS coordinates, with each station ID registered as a dedicated vocabulary token so routes cannot be hallucinated via character composition.

Songhao Wu, Zhongxin Chen, Yuxuan Liu, Heng Cui, Cong Li, Rui Yan

The paper introduces EmbedFilter, a linear post-processing transformation that improves text embeddings extracted zero-shot from LLMs. It works by taking the SVD of the LLM's unembedding matrix and projecting embeddings onto only the mid-range ('bulk') singular vectors, discarding the 'edge spectrum' (largest and smallest singular values) that the authors show encodes high-frequency, semantically empty tokens.