AI Papers of the Week
Every paper worth reading in AI, hand-picked one week at a time.
Retrieval-Augmented Reasoning Model
Introduces RARE, a new paradigm for training domain-specific LLMs that focuses on reasoning, not memorization. Key ideas: ● Inspired by Bloom’s Taxonomy – RARE shifts LLM training from memorizing knowledge (“Remember”) to applying and evaluating it (“Analyze”, “Create”). It separates domain knowledge (retrieved externally) from domain thinking (learned during training), enabling better performance under tight parameter budgets. ● Open-book prepared training – RARE injects retrieved knowledge into training prompts, letting models learn reasoning patterns instead of rote facts. This open-book, reasoning-first setup beats both standard SFT and RAG approaches, especially in medicine. ● Massive accuracy gains with small models – On five medical QA benchmarks, RARE-trained Llama-3.1-8B and Qwen-2.5-7B outperformed GPT-4 + RAG, with up to +20% accuracy boosts (e.g., PubMedQA: 78.63% vs. GPT-4’s 75.2%, CoVERT: 74.14% vs. GPT-4’s 65.67%). ● Training via distillation + adaptive retries – RARE distills answers (and reasoning paths) from a strong teacher (e.g., QwQ-32B), refining outputs until a correct answer is found. This creates a high-quality dataset that teaches contextualized, case-based thinking. ● New role for retrieval – Unlike standard RAG (used only at inference), RARE uses retrieval during training to shape reasoning. It models knowledge integration (p(kx, R(x))) and reasoning (p(rx, R(x), k)) as separate steps, replacing memorization with application. Overall, this work reframes LLM training for domain-specific intelligence: externalize facts, internalize reasoning. It unlocks strong performance from small models without overfitting or hallucination.
Why do LLMs Attend to First Token?
This new paper explains why LLMs obsessively focus attention on the first token — a phenomenon known as an attention sink. Their theory: it’s a useful trick to prevent representational collapse in deep Transformers. ● Sinks = over-mixing shields – LLMs with long contexts and deep layers tend to over-mix information, causing similar embeddings for all tokens (i.e., rank collapse or over-squashing). Attention sinks—where many heads fixate on the ⟨bos⟩ token—act as no-ops that reduce token interaction and preserve representation diversity across layers. ● Sharp experiments on Gemma & LLaMa – Perturbation tests in Gemma 7B show ⟨bos⟩ significantly slows the spread of changes through the model. Meanwhile, in LLaMa 3.1 models, over 80% of attention heads show strong sink behavior in the 405B variant, supporting the theory that larger models need stronger sinks. ● Sinks emerge naturally – Even without special pretraining, sinks tend to form at the first position, not because of the ⟨bos⟩ token itself, but due to its location. However, if ⟨bos⟩ is fixed during training and later removed, performance collapses, showing that sink formation is data-dependent. ● Theoretical grounding – The authors connect sink emergence to Jacobian norm bounds, proving that sinks reduce sensitivity to token perturbations. Their math shows that deeper models and longer contexts require stronger sinks. ● Layerwise dynamics insight – Some attention heads use ⟨bos⟩ as a “default” target, unless a special pattern (e.g., apostrophe) triggers real computation. This supports a conditional attention mechanism—attend to ⟨bos⟩ unless needed elsewhere.
Efficient Test-time Scaling with Code
Z1 is a new method for making large language models more compute-efficient at test time, especially during reasoning. The core idea is to train LLMs with short and long code-based reasoning trajectories, and then dynamically adjust reasoning depth during inference. Key contributions: ● Z1-Code-Reasoning-107K dataset – They construct a 107K-sample dataset with short and long reasoning paths for simple and complex coding problems. Trajectories are distilled from QwQ-32B and paired to help the model learn when to stop thinking. ● Shifted Thinking Window – A new test-time strategy that eliminates explicit <think delimiters. Instead, the model adapts reasoning token budget based on problem difficulty. Simple problems invoke shallow reasoning; complex ones get capped (e.g., 4096 tokens max), with hints nudging the model to finalize the answer. ● Big efficiency gains – The 7B-scale model Z1-7B matches R1-Distill-Qwen-7B across multiple reasoning tasks (MATH500, LiveCodeBench, GPQA Diamond) but with ~30% of the reasoning tokens. For instance, on GPQA Diamond, Z1-7B achieves 47.5% while using less than half the tokens. ● Code reasoning transfers to general tasks – Despite being trained only on code-based CoT data, Z1 generalizes well to broader domains like science and math, outperforming other 7B reasoning models (e.g., OpenThinker-7B, s1.1-7B) across multiple benchmarks. ● What makes reasoning data effective? – Ablation studies reveal two key dataset design levers: (1) longer reasoning trajectories improve inference quality; (2) larger training sample sizes boost average thinking time and accuracy, even without altering trajectory length.
A Survey of Efficient Reasoning for LLMs
This survey focuses on reasoning economy in LLMs, analyzing how to balance deep reasoning performance with computational cost. It reviews inefficiencies, behavioral patterns, and potential solutions at both post-training and inference stages.
Hidden Factual Knowledge in LLMs
This study introduces a framework to measure hidden knowledge in LLMs, showing that models encode significantly more factual information internally than they express in outputs, up to 40% more. It also finds that some answers, although known internally, are never generated, highlighting key limits in test-time sampling for QA tasks.
Qwen2.5-Omni
Qwen2.5-Omni is a single end-to-end multimodal model that can perceive and understand text, audio, image, and video, and generate both text and speech in real time. It introduces architectural and training innovations that push the boundaries of streaming, multi-signal intelligence. Highlights: ● Thinker-Talker architecture – Inspired by the human brain and mouth, Qwen2.5-Omni separates reasoning (Thinker) and speech generation (Talker). Thinker (a transformer decoder) handles all perception and text generation. Talker (a dual-track autoregressive decoder) generates speech by consuming both text and hidden states from Thinker. Together, they’re trained end-to-end for synchronized text-speech output. ● Streaming-first design – To support real-time interaction, Qwen2.5-Omni implements block-wise encoders (for audio and vision) and a sliding-window codec generator for streaming audio. The model introduces TMRoPE (Time-aligned Multimodal RoPE), a 3D positional encoding system that aligns video and audio inputs to the same time axis. ● Pretraining scale & alignment – Trained on over 1.2 trillion tokens of diverse multimodal data, including 300B audio and 100B video-audio tokens. Uses instruction-tuned ChatML formatting and performs multi-stage post-training for both Thinker and Talker. Talker undergoes RL fine-tuning (DPO) and multi-speaker adaptation to ensure natural, stable speech output. ● SOTA across modalities – Qwen2.5-Omni achieves state-of-the-art on OmniBench, surpasses Qwen2-Audio in ASR/S2TT, and matches or beats Qwen2.5-VL in image and video tasks. On SEED zero-shot TTS, it outperforms CosyVoice 2 and F5-TTS in naturalness and stability, with low WER and high speaker similarity. ● Closes the voice-text gap – On a voice-instruction benchmark (converted from MMLU, GSM8K, etc.), Qwen2.5-Omni nearly matches its own text-instructed sibling Qwen2-7B, showing dramatic improvements in speech-based instruction following.
Chain-of-Tools
This new paper presents Chain-of-Tools (CoTools), a new method to enable LLMs to incorporate expansive external toolsets—including tools never seen during training—while preserving CoT (chain-of-thought) reasoning. Highlights: ● Frozen LLM with lightweight fine-tuning – Unlike conventional approaches, CoTools keeps the LLM’s parameters frozen, instead fine-tuning separate modules (a Tool Judge and Tool Retriever) on top of the model’s hidden states. This preserves the LLM’s core capabilities while letting it call an open-ended set of tools during reasoning. ● Massive unseen tools – CoTools treats tools as semantic vectors computed from their textual descriptions. Even tools that never appear in the fine-tuning data can be invoked if they match the model’s query vectors, enabling new tools to be plugged in without retraining the entire system. ● Tool calls integrated into CoT – The system determines whether and when to call a tool in the middle of generating an answer. It then selects the best tool from thousands of candidates based on learned representations of the query and partial solution context. This helps to significantly boost accuracy on complex tasks. ● Strong gains on reasoning and QA – Experiments on GSM8K-XL, FuncQA, KAMEL, and the newly introduced SimpleToolQuestions dataset (with 1,836 tools) show improved tool-selection accuracy and superior final answers versus baseline methods. Notably, CoTools consistently scales to large tool pools and generalizes to unseen tools.
A Review of DeepSeek Models
This paper provides an in-depth review of the cutting-edge techniques behind DeepSeek's open-source LLMs—DeepSeek-V3 and DeepSeek-R1. These models achieve state-of-the-art performance with significantly lower resource requirements compared to proprietary counterparts. Key highlights include:
Towards Hierarchical Multi-Step Reward Models for Enhanced Reasoning in LLMs
It proposes a Hierarchical Reward Model (HRM) that addresses reward hacking and error propagation issues in fine-grained LLM reasoning. They also introduce Hierarchical Node Compression (HNC) to augment MCTS-based automatic data annotation, boosting label diversity and robustness at minimal computational cost.
Compute Optimal Scaling of Skills
Researchers from the University of Wisconsin and Meta AI investigate how different skills (knowledge-based QA vs. code generation) exhibit contrasting optimal scaling behaviors in LLMs. Their key question: does the compute-optimal trade-off between model size and data volume depend on the type of skill being learned? Surprisingly, the answer is yes—they show distinct “data-hungry” vs. “capacity-hungry” preferences per skill. Highlights:
DeepMesh
Researchers from Tsinghua University, Nanyang Technological University, and ShengShu propose DeepMesh, a transformer-based system that generates high-quality 3D meshes with artist-like topology. Key ideas include:
Gemma 3
Gemma 3 is a lightweight open model family (1B–27B parameters) that integrates vision understanding, multilingual coverage, and extended context windows (up to 128K tokens). Here is everything you need to know:
Transformers without Normalization
Researchers from Meta, NYU, MIT, and Princeton present a surprisingly simple method, Dynamic Tanh (DyT), that removes normalization layers (e.g. LayerNorm, RMSNorm) in Transformers while achieving equal or better results. Key ideas include:
Auditing LLMs for Hidden Objectives
Anthropic proposes a new framework for systematically auditing LLMs to uncover hidden goals or objectives that go beyond what users and developers explicitly intend. The researchers deliberately train a language model with a concealed objective (making it exploit reward model flaws in RLHF) and then attempt to expose it with different auditing techniques.
Post Training of LLMs
PoLMs like OpenAI-o1/o3 and DeepSeek-R1 tackle LLM shortcomings in reasoning, ethics, and specialized tasks. This survey tracks their evolution and provides a taxonomy of techniques across fine-tuning, alignment, reasoning, efficiency, and integration, guiding progress toward more robust, versatile AI.
A Few Tokens Are All You Need
Researchers from Tencent AI Lab and The Chinese University of Hong Kong, Shenzhen propose a new approach to boost reasoning in LLMs by only fine-tuning on the first few tokens of generated solutions. Key ideas include:
A Deep Dive into Reasoning LLMs
This survey explores how LLMs can be enhanced after pretraining through fine-tuning, reinforcement learning, and efficient inference strategies. It also highlights challenges like catastrophic forgetting, reward hacking, and ethical considerations, offering a roadmap for more capable and trustworthy AI systems.
Cognitive Behaviors that Enable Self-Improving Reasoners
Researchers from Stanford University and colleagues investigate why some language models excel in reinforcement learning (RL)-based self-improvement, while others quickly plateau. The study identifies four cognitive behaviors-verification, backtracking, subgoal setting, and backward chaining-that underpin successful problem-solving in both humans and language models. Key findings:
Conversational Speech Model
Researchers from Sesame propose an end-to-end multimodal TTS approach for natural, context-aware speech in real-time conversational AI systems.
Fractal Generative Models
Researchers from MIT CSAIL & Google DeepMind introduce a novel fractal-based framework for generative modeling, where entire generative modules are treated as atomic "building blocks" and invoked recursively-resulting in self-similar fractal architectures:
GPT-4.5
OpenAI introduces GPT-4.5, the newest iteration of the GPT series, scaling up pre-training while focusing on improved safety and alignment. Key insights include:
Emergent Misalignment
New research investigates an unexpected phenomenon: finetuning an LLM on a narrow task can cause it to become broadly misaligned across unrelated domains. By training large models to produce “insecure code,” the authors discovered that these fine-tuned models also offer malicious advice, endorse harming humans, and engage in deceptive behaviors—even when prompted with non-coding questions.
A Systematic Survey of Prompt Optimization
This paper offers a comprehensive survey of Automatic Prompt Optimization (APO)—defining its scope, presenting a unifying 5-part framework, categorizing existing methods, and highlighting key progress and challenges in automating prompt engineering for LLMs.
The AI CUDA Engineer
Sakana AI introduces The AI CUDA Engineer, an end-to-end agentic system that can produce highly optimized CUDA kernels. Key contributions: