#llm

23 Dec 2025 · Paper summary
Paper summary: DataFlow: An LLM-Driven Framework for Unified Data Preparation and Workflow Automation in the Era of Data-Centric AI
DataFlow is an LLM-driven framework for building data-preparation pipelines from about 200 reusable operators with a PyTorch-style API, plus an agent that turns plain-language specs into pipelines.
2 Dec 2025 · Paper summary
Paper summary: From Code Foundation Models to Agents and Applications: A Comprehensive Survey and Practical Guide to Code Intelligence
A large survey and practical guide to code LLMs, from data curation and pre-training through SFT, RL and coding agents, comparing general and code-specialised models and the gap between benchmarks and real development.
30 Oct 2025 · Paper summary
Paper summary: Scaling Latent Reasoning via Looped Language Models
Ouro builds latent reasoning into pre-training with looped language models (iterating in latent space with a learned depth), and its 1.4B and 2.6B models are reported to match up to 12B models.
20 Oct 2025 · Paper summary
Paper summary: A Theoretical Study on Bridging Internal Probability and Self-Consistency for LLM Reasoning
A theoretical framework for sampling-based test-time scaling shows self-consistency has high estimation error and perplexity has high modelling error; the hybrid RPC method matches self-consistency with half the samples.
14 Oct 2025 · Paper summary
Paper summary: QeRL: Beyond Efficiency -- Quantization-enhanced Reinforcement Learning for LLMs
QeRL combines NVFP4 quantisation with LoRA for RL training of LLMs; beyond saving memory and speeding up rollouts, quantisation noise raises policy entropy and helps exploration. It enables RL on a 32B model with one H100.
1 Oct 2025 · Paper summary
Paper summary: The Dragon Hatchling: The Missing Link between the Transformer and Models of the Brain
The Dragon Hatchling (BDH) is a brain-inspired language-model architecture of locally interacting "neuron particles" with Hebbian working memory, reported to match GPT-2 at the same size and to be more interpretable.
23 Sept 2025 · Paper summary
Paper summary: Qwen3-Omni Technical Report
Qwen3-Omni is a single model across text, image, audio and video that's reported to match same-sized single-modal Qwen models, with a Thinker–Talker MoE design and 234 ms theoretical first-packet speech latency.
19 Sept 2025 · Paper summary
Paper summary: FlowRL: Matching Reward Distributions for LLM Reasoning
FlowRL replaces reward maximisation with matching the whole reward distribution (via a learnable partition function and reverse KL), to keep diverse reasoning paths alive; it reports +10.0% over GRPO on maths.
10 Sept 2025 · Paper summary
Paper summary: Sharing is Caring: Efficient LM Post-Training with Collective RL Experience Sharing
SAPO is a fully decentralised, asynchronous RL post-training method in which many heterogeneous nodes train their own models but share rollouts, reporting up to 94% higher cumulative reward in controlled tests.
1 Sept 2025 · Paper summary
Paper summary: A.S.E: A Repository-Level Benchmark for Evaluating Security in AI-Generated Code
A.S.E is a repository-level benchmark for the security of AI-generated code; it finds that current LLMs still struggle with secure coding, and that a bigger reasoning budget doesn't reliably help.
11 Aug 2025 · Paper summary
Paper summary: GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models
GLM-4.5 is an open MoE LLM (355B total, 32B active) with hybrid thinking and direct modes, trained on 23T tokens plus expert iteration and RL, reporting 70.1% TAU-Bench, 91.0% AIME 24 and 64.2% SWE-bench Verified.
25 Jul 2025 · Paper summary
Paper summary: Group Sequence Policy Optimization
Group Sequence Policy Optimization computes importance ratios and clipping at the level of whole sequences rather than tokens, and is reported to be more stable and efficient than GRPO, especially for MoE models.
18 Jul 2025 · Paper summary
Paper summary: A Survey of Context Engineering for Large Language Models
A survey of over 1,400 papers that defines "context engineering": retrieving, processing and managing everything put into an LLM's context, and how RAG, memory, tools and multi-agent systems build on it.
8 Jul 2025 · Paper summary
Paper summary: MemOS: A Memory OS for AI System
MemOS proposes a memory operating system for LLMs that manages plaintext, activation-based and parameter-level memory as one schedulable resource, built around "MemCube" units with provenance and versioning.
23 Jun 2025 · Paper summary
Paper summary: Drag-and-Drop LLMs: Zero-Shot Prompt-to-Weights
Drag-and-Drop LLMs train a generator that maps a few unlabelled task prompts straight to LoRA weights, specialising an LLM for a new task in seconds without any gradient-based fine-tuning on it.
17 Jun 2025 · Paper summary
Paper summary: MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention
MiniMax-M1 is an open-weight hybrid-attention reasoning model (456B total, 45.9B active, 1M-token context) trained with large-scale RL and a new algorithm, CISPO, at a reported rental cost of $534,700.
10 Jun 2025 · Paper summary
Paper summary: Reinforcement Pre-Training
Reinforcement Pre-Training reframes next-token prediction as a reasoning task: the model reasons before predicting the next token and gets a verifiable reward when it's right, so ordinary text becomes RL training data.
4 Jun 2025 · Paper summary
Paper summary: Reflect, Retry, Reward: Self-Improving LLMs via Reinforcement Learning
When a model fails a task, it writes a self-reflection and tries again; if the retry succeeds, the reflection tokens are rewarded, which trains better self-reflection with only binary feedback.
19 May 2025 · Paper summary
Paper summary: Qwen3 Technical Report
Qwen3 merges a "thinking" mode for multi-step reasoning and a fast non-thinking mode in one model family (0.6B to 235B, dense and MoE), adds a thinking budget, and expands to 119 languages.
7 May 2025 · Paper summary
Paper summary: Absolute Zero: Reinforced Self-play Reasoning with Zero Data
Absolute Zero trains a model to propose its own coding and maths-style tasks and then solve them, with a code executor checking both, so reasoning improves through self-play with no external training data.
30 Apr 2025 · Paper summary
Paper summary: Reinforcement Learning for Reasoning in Large Language Models with One Training Example
RL with verifiable rewards on a single maths training example lifts Qwen2.5-Math-1.5B on MATH500 from 36.0% to 73.6%, matching training on a 1.2k-example dataset.
21 Apr 2025 · Paper summary
Paper summary: Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?
Measuring pass@k at large k, this study finds RL with verifiable rewards makes models better at k = 1 but doesn't expand what the base model can solve at all; distillation, by contrast, can.
27 Mar 2025 · Paper summary
Paper summary: Qwen2.5-Omni Technical Report
Qwen2.5-Omni takes text, images, audio and video as input and streams back both text and speech, using a Thinker (LLM) and a Talker (speech generator) that share hidden states.
11 Mar 2025 · Paper summary
Paper summary: Feature-Level Insights into Artificial Text Detection with Sparse Autoencoders
Using sparse-autoencoder features from Gemma-2-2B's residual stream to detect and explain AI-generated text, revealing a distinct LLM writing style, especially in information-dense domains.
7 Mar 2025 · Paper summary
Paper summary: START: Self-taught Reasoner with Tools
START teaches a long chain-of-thought reasoning model to call a Python interpreter mid-reasoning, using inserted hints ("Wait, maybe using Python here is a good idea") and rejection-sampling fine-tuning.
24 Feb 2025 · Paper summary
Paper summary: LLM-Microscope: Uncovering the Hidden Role of Punctuation in Context Memory of Transformers
Tokens we treat as filler (punctuation, articles, stop words) carry a surprising share of an LLM's contextual memory; removing them hurts MMLU and long-context scores, and the paper releases a toolkit to measure this.
10 Feb 2025 · Paper summary
Paper summary: Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach
A language model that "thinks" by looping a recurrent block more times at test time, reasoning in latent space instead of writing out more chain-of-thought tokens; a 3.5B proof of concept.
6 Feb 2025 · Paper summary
Paper summary: SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model
SmolLM2 is a 1.7B model "overtrained" on about 11 trillion tokens in stages that mix web, maths, code and instruction data, with new datasets (FineMath, Stack-Edu, SmolTalk) built where existing ones were too small or noisy.
29 Jan 2025 · Paper summary
Paper summary: SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training
On a card-arithmetic game and a real-world navigation task, models post-trained with RL generalise to unseen rule and visual variants, while SFT tends to memorise the training data.
23 Jan 2025 · Paper summary
Paper summary: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
DeepSeek-R1 shows that large-scale reinforcement learning, without human-labelled reasoning traces, can bring out advanced reasoning such as self-reflection and verification in a pretrained LLM.
15 Jan 2025 · Paper summary
Paper summary: MiniMax-01: Scaling Foundation Models with Lightning Attention
MiniMax-01 combines lightning attention, a linear-attention variant, with a 32-expert MoE (456B total, 45.9B active) to train on 1M-token contexts and extrapolate to 4M at inference.
9 Jan 2025 · Paper summary
Paper summary: rStar-Math: Small LLMs Can Master Math Reasoning with Self-Evolved Deep Thinking
rStar-Math gets small language models to strong maths reasoning with Monte Carlo Tree Search guided by a small process preference model, both improved through self-evolution without distillation from stronger models.
27 Dec 2024 · Paper summary
Paper summary: YuLan-Mini: An Open Data-efficient Language Model
YuLan-Mini is a 2.42B-parameter base model trained on 1.08T tokens that the authors report matches models trained on much more data, with full details of the data and training released.
20 Dec 2024 · Paper summary
Paper summary: Qwen2.5 Technical Report
Qwen2.5 scales pre-training data from 7 to 18 trillion tokens and uses over a million SFT samples plus multistage RL, across a wide range of open-weight sizes.
8 Nov 2024 · Paper summary
Paper summary: OpenCoder: The Open Cookbook for Top-Tier Code Large Language Models
OpenCoder is a top-tier code LLM released as an "open cookbook": weights and inference code plus reproducible training data, the full data-processing pipeline and training protocols.
8 Oct 2024 · Paper summary
Paper summary: Differential Transformer
Differential attention computes attention as the difference between two softmax maps, cancelling common-mode noise; Diff Transformer is reported to beat Transformers on scaling, long context, retrieval and hallucination.
30 Sept 2024 · Paper summary
Paper summary: Emu3: Next-Token Prediction is All You Need
Emu3 tokenises images, text and video into one discrete space and trains a single Transformer from scratch with only next-token prediction, reported to beat SDXL at generation and LLaVA-1.6 at perception.
19 Sept 2024 · Paper summary
Paper summary: Qwen2.5-Coder Technical Report
Qwen2.5-Coder is a family of six code models (0.5B to 32B) continued-pretrained on over 5.5 trillion tokens, reported to set state-of-the-art results on more than 10 code benchmarks for their sizes.
10 Sept 2024 · Paper summary
Paper summary: Towards a Unified View of Preference Learning for Large Language Models: A Survey
A survey that breaks every LLM preference-alignment method into four components, model, data, feedback and algorithm, to show how RLHF, DPO and their relatives connect.
28 Aug 2024 · Paper summary
Paper summary: Writing in the Margins: Better Inference Pattern for Long Context Retrieval
Writing in the Margins processes a long context chunk by chunk during prefill, writes short "margin" notes per chunk, keeps the relevant ones, and uses them to answer, improving off-the-shelf models without fine-tuning.
13 Aug 2024 · Paper summary
Paper summary: The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery
The AI Scientist chains LLM steps that generate research ideas, write and run experiment code, write a full paper and review it with an automated reviewer, as a framework for fully automated research.
16 Jul 2024 · Paper summary
Paper summary: Qwen2 Technical Report
The Qwen2 report covers open dense models from 0.5B to 72B plus an MoE, with Qwen2-72B reaching 84.2 on MMLU as a base model and support for about 30 languages.
1 Jul 2024 · Paper summary
Paper summary: Scaling Synthetic Data Creation with 1,000,000,000 Personas
Persona Hub collects one billion personas from web data and uses them to steer an LLM toward diverse viewpoints when generating synthetic data: maths problems, instructions, knowledge texts, game NPCs and tools.
26 Jun 2024 · Paper summary
Paper summary: The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale
FineWeb is a 15-trillion-token open pre-training dataset built from 96 Common Crawl snapshots, with every curation choice ablated; its 1.3T-token educational subset, FineWeb-Edu, gives large gains on MMLU and ARC.
21 Jun 2024 · Paper summary
Paper summary: Instruction Pre-Training: Language Models are Supervised Multitask Learners
Instruction Pre-Training augments raw pre-training text with synthesised instruction–response pairs (200M of them across 40+ task types) and reports stronger base models and better gains from later instruction tuning.
17 May 2024 · Paper summary
Paper summary: Chameleon: Mixed-Modal Early-Fusion Foundation Models
Chameleon is one token-based model that reads and generates images and text in any order, using early fusion, plus a training recipe designed to keep this mixed-modal setup stable.
3 May 2024 · Paper summary
Paper summary: Prometheus 2: An Open Source Language Model Specialized in Evaluating Other Language Models
Prometheus 2 is an open evaluator LM that does both direct scoring and pairwise ranking against user-defined criteria, built to track human and GPT-4 judgements more closely than earlier open judges.
23 Apr 2024 · Paper summary
Paper summary: Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone
phi-3-mini is a 3.8B model trained on 3.3T tokens of heavily filtered web and synthetic data, reported to reach 69% on MMLU while being small enough to run on a phone.
16 Apr 2024 · Paper summary
Paper summary: Learn Your Reference Model for Real Good Alignment
Trust Region alignment (TR-DPO, TR-IPO, TR-KTO) updates the reference policy during offline alignment training, which reduces over-optimisation.
11 Apr 2024 · Paper summary
Paper summary: Leave No Context Behind: Efficient Infinite Context Transformers with Infini-attention
Infini-attention adds a compressive memory to standard attention, combining local masked attention with long-term linear attention so Transformers can stream unbounded inputs with bounded memory.
1 Apr 2024 · Paper summary
Paper summary: Jamba: A Hybrid Transformer-Mamba Language Model
Jamba interleaves Transformer and Mamba layers and adds mixture-of-experts in some layers; the released configuration fits on one 80GB GPU and handles contexts up to 256K tokens.
27 Mar 2024 · Paper summary
Paper summary: The Unreasonable Ineffectiveness of the Deeper Layers
Removing up to about half of the layers of some open LLMs, chosen by layer similarity and followed by a little QLoRA "healing", barely hurts question-answering performance.
21 Mar 2024 · Paper summary
Paper summary: LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models
LlamaFactory is an open framework that brings many efficient fine-tuning methods together for 100+ LLMs, including a no-code web UI called LlamaBoard.
15 Mar 2024 · Paper summary
Paper summary: MM1: Methods, Analysis & Insights from Multimodal LLM Pre-training
Apple's MM1 ablates how to build multimodal LLMs and finds the data mix and image encoder (resolution, token count) matter a lot, while the vision–language connector design matters little.
7 Mar 2024 · Paper summary
Paper summary: GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection
GaLore projects gradients into a low-rank subspace for the optimizer state while still updating all weights, cutting optimizer memory by up to 65.5% (82.5% in 8-bit) without LoRA's accuracy trade-off.
28 Feb 2024 · Paper summary
Paper summary: The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits
BitNet b1.58 restricts every weight to −1, 0 or +1 (about 1.58 bits) and reports matching full-precision LLMs of the same size and data on perplexity and end tasks, at much lower cost.
22 Feb 2024 · Paper summary
Paper summary: LongRoPE: Extending LLM Context Window Beyond 2 Million Tokens
LongRoPE extends a pretrained LLM's context window to 2048k tokens by searching for non-uniform RoPE interpolation, extending progressively, and readjusting for short contexts.
16 Feb 2024 · Paper summary
Paper summary: Chain-of-Thought Reasoning Without Prompting
Chain-of-thought reasoning can appear without any prompting: branching on the top-k first tokens during decoding often reveals reasoning paths, and those paths come with higher answer confidence.
6 Feb 2024 · Paper summary
Paper summary: DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models
DeepSeekMath 7B continues pre-training on 120B maths tokens mined from Common Crawl and introduces GRPO, a PPO variant without a value network, reaching 51.7% on the MATH benchmark.
2 Feb 2024 · Paper summary
Paper summary: OLMo: Accelerating the Science of Language Models
OLMo is a competitive language model released with its training data, training code and evaluation code, not just weights, so others can study how such models are built.
18 Jan 2024 · Paper summary
Paper summary: Self-Rewarding Language Models
A model judges its own responses with LLM-as-a-Judge prompting and trains on those rewards with iterative DPO, so both its instruction following and its judging improve over three rounds.
9 Jan 2024 · Paper summary
Paper summary: Mixtral of Experts
Mixtral 8x7B is a sparse mixture-of-experts model: each layer has 8 feed-forward experts and a router picks 2 per token, giving 47B total but 13B active parameters.
27 Dec 2023 · Paper summary
Paper summary: SOLAR 10.7B: Scaling Large Language Models with Simple yet Effective Depth Up-Scaling
SOLAR 10.7B grows a smaller LLM by duplicating its layer stack, trimming the overlap and continuing pre-training ("depth up-scaling"), a simpler route than mixture-of-experts.
20 Dec 2023 · Paper summary
Paper summary: LLM in a flash: Efficient Large Language Model Inference with Limited Memory
Apple researchers run LLMs larger than the available DRAM by keeping weights in flash and loading only what is needed, with windowing and row–column bundling designed around how flash memory reads data.
12 Dec 2023 · Paper summary
Paper summary: LLM360: Towards Fully Transparent Open-Source LLMs
LLM360 argues for releasing everything behind an LLM (data, code, intermediate checkpoints, analyses) and starts with two fully open 7B models, Amber and CrystalCoder.
4 Dec 2023 · Paper summary
Paper summary: Mamba: Linear-Time Sequence Modeling with Selective State Spaces
Mamba makes state-space model parameters depend on the input, so the model can choose what to remember or forget, and pairs this with a hardware-aware scan: linear-time sequence modelling without attention.
3 Dec 2023 · Paper summary
Paper summary: Merlin:Empowering Multimodal LLMs with Foresight Minds
Merlin trains a multimodal LLM to predict object trajectories and then reason about future events, using two stages called Foresight Pre-Training and Foresight Instruction-Tuning.
30 Oct 2023 · Paper summary
Paper summary: CodeFusion: A Pre-trained Diffusion Model for Code Generation
CodeFusion generates code with a diffusion model that iteratively denoises a whole program conditioned on the natural-language request, instead of writing tokens left to right.
18 Oct 2023 · Paper summary
Paper summary: BitNet: Scaling 1-bit Transformers for Large Language Models
BitNet replaces nn.Linear with BitLinear and trains Transformers with 1-bit weights from scratch, reporting competitive perplexity with much lower memory and energy than FP16 or 8-bit baselines.
12 Oct 2023 · Paper summary
Paper summary: Prometheus: Inducing Fine-grained Evaluation Capability in Language Models
Prometheus is an open 13B evaluator LLM trained on GPT-4-generated feedback that scores long-form answers against a user-provided rubric, approaching GPT-4's agreement with human raters.
22 Sept 2023 · Paper summary
Paper summary: LongLoRA: Efficient Fine-tuning of Long-Context Large Language Models
LongLoRA extends the context window of pretrained LLMs cheaply by fine-tuning with shifted sparse local attention plus a LoRA variant that also trains embeddings and norms.
12 Sept 2023 · Paper summary
Paper summary: Textbooks Are All You Need II: phi-1.5 technical report
phi-1.5 is a 1.3B-parameter model trained mostly on synthetic "textbook-quality" data, reported to match models about five times larger on common-sense reasoning.
8 Sept 2023 · Paper summary
Paper summary: Large Language Models as Optimizers
OPRO describes an optimisation task in natural language and lets an LLM propose new solutions from a history of past solutions and scores, used mainly to optimise prompts.
14 Aug 2023 · Paper summary
Paper summary: Self-Alignment with Instruction Backtranslation
Instruction backtranslation lets a language model label unlabelled web text with the instructions it answers, then pick the best pairs to fine-tune itself.