AI Papers of the Week
Every paper worth reading in AI, hand-picked one week at a time.
Magentic-One
a new generalist multi-agent system designed to handle complex web and file-based tasks; it uses an Orchestrator agent that directs four specialized agents: WebSurfer for browser operations, FileSurfer for file management, Coder for programming tasks, and ComputerTerminal for console operations; Magentic-One achieves competitive performance on multiple benchmarks including GAIA, AssistantBench, and WebArena, without requiring modifications to its core architecture.
Mixtures of In-Context Learners
uses subsets of demonstrations to train experts via in-context learning; given a training set, a trainable weighting function is used to combine the experts' next-token predictions; this approach applies to black-box LLMs since access to the internal parameters of the LLM is not required. Good properties include the following: 1) competitive with standard ICL while being significantly more data, memory, and computationally efficient, and 2) resilient to noisy demonstrations and label imbalance.
Attacking Vision-Language Agents via Pop-ups
shows that integrating adversarial pop-ups into existing agent testing environments leads to an attack success rate of 86%; this decreases the agents' task success rate by 47%; they also add that basic defense techniques (e.g., instructing the agent to ignore pop-ups) are ineffective.
Multi-expert Prompting with LLMs
improves LLM responses by simulating multiple experts and aggregating their responses; it guides an LLM to fulfill input instructions by simulating multiple experts and selecting the best response among individual and aggregated views; it achieves a new state-of-the-art on TruthfulQA-Generation with ChatGPT, surpassing the current SOTA of 87.97%; it also improves performance across factuality and usefulness while reducing toxicity and hurtfulness.
Number Understanding of LLMs
provides a comprehensive analysis of the numerical understanding and processing ability (NUPA) of LLMs; finds that naive finetuning can improve NUPA a lot on many but not all tasks; it also reports that techniques designed to enhance NUPA prove ineffective for finetuning pretrained models; explores chain-of-thought techniques applied to NUPA and suggests that chain-of-thought methods face scalability challenges, making them difficult to apply in practical scenarios.
WebRL
proposes a self-evolving online curriculum RL framework to bridge the gap between open and proprietary LLM-based web agents; it improves the success rate of Llama-3.1-8B from 4.8% to 42.4%, and from 6.1% to 43% for GLM4-9B; the open models significantly surpass the performance of GPT-4-Turbo (17.6%) and GPT-4o (13.9%); the self-evolving curriculum addresses the scarcity of web agent training tasks; this is underpinned by a robust outcome-supervised reward model to evaluate task success; an adaptive RL strategy helps to deal with distribution drift in online learning and ensures consistent improvements.
Adapting while Learning
proposes a two-part fine-tuning approach that first helps LLMs learn from tool-generated solutions and then trains them to determine when to solve problems directly versus when to use tools; testing on math, climate science, and epidemiology benchmarks shows significant improvements, with a 28% boost in accuracy and 14% better tool usage precision compared to leading models like GPT-4 and Claude-3.5; the two-stage approach helps the LLM to adaptively solve scientific problems of varying complexity.
Personalization of LLMs
presents a comprehensive framework for understanding personalized LLMs; introduces taxonomies for different aspects of personalization and unifying existing research across personalized text generation and downstream applications.
Geometry of Concepts in LLMs
examines the geometric structure of concept representations in sparse autoencoders (SAEs) at three scales: 1) atomic-level parallelogram patterns between related concepts (e.g., man:woman::king:queen), 2) brain-like functional "lobes" for different types of knowledge like math/code, 3) and galaxy-level eigenvalue distributions showing a specialized structure in middle model layers.
SimpleQA
a challenging benchmark of 4,326 short factual questions adversarially collected against GPT-4 responses; reports that frontier models like GPT-4o and Claude achieve less than 50% accuracy; finds that there is a positive calibration between the model stated confidence and accuracy, signaling that they have some notion of confidence; claims that there is still room to improve the calibration of LLMs in terms of stated confidence.
Automating Agentic Workflow Generation
a novel framework for automating the generation of agentic workflows; it reformulates workflow optimization as a search problem over code-represented workflows, where edges connect LLM-invoking nodes; it efficiently explores the search space using a variant of MCTS, iteratively refining workflows through code modification, tree-structured experience, and execution feedback; experiments across six benchmark datasets demonstrate AFlow’s effectiveness, showing a 5.7% improvement over manually designed methods and a 19.5% improvement over existing automated approaches; AFlow also enables smaller models to outperform GPT-4o on specific tasks at just 4.55% of its inference cost.
LLMs Solve Math with a Bag of Heuristics
uses causal analysis to find neurons that explain an LLM's behavior when doing basic arithmetic logic; discovers and hypothesizes that the combination of heuristic neurons is the mechanism used to produce correct arithmetic answers; finds that the unordered combination of different heuristic types is the mechanism that explains most of the model’s accuracy on arithmetic prompts.
o1 Replication Journey
reports to be replicating the capabilities of OpenAI's o1 model; their journey learning technique encourages learning not just shortcuts, but the complete exploration process, including trial and error, reflection, and backtracking; claims that with only 327 training samples, their journey learning technique surpassed shortcut learning by 8.0% on the MATH dataset.
Distinguishing Ignorance from Error in LLM Hallucinations
a method to distinguish between two types of LLM hallucinations: when models lack knowledge (HK-) versus when they hallucinate despite having correct knowledge (HK+); they build model-specific datasets using their proposed approach and show that model-specific datasets are more effective for detecting HK+ hallucinations compared to generic datasets.
Multimodal RAG
provides a discussion on how to best integrate multimodal models into RAG systems for the industrial domain; it also provides a deep discussion on the evaluation of these systems using LLM-as-a-Judge.
The Role of Prompting and External Tools in Hallucination Rates of LLMs
tests different prompting strategies and frameworks aimed at reducing hallucinations in LLMs; finds that simpler prompting techniques outperform more complex methods; it reports that LLM agents exhibit higher hallucination rates due to the added complexity of tool usage.
MrT5
a more efficient variant of byte-level language models that uses a dynamic token deletion mechanism (via a learned delete gate) to shorten sequence lengths by up to 80% while maintaining model performance; this enables faster inference and better handling of multilingual text without traditional tokenization; MrT5 maintains competitive accuracy with ByT5 on downstream tasks such as XNLI and character-level manipulations while improving inference runtimes.
Relaxed Recursive Transformers
introduces a novel approach, Relaxed Recursive Transformer, that significantly reduces LLM size through parameter sharing across layers while maintaining performance; the model is initialized from standard pretrained Transformers, but only uses a single block of unique layers that is repeated multiple times in a loop; then it adds flexibility to the layer tying constraint via depth-wise low-rank adaptation (LoRA) modules; shows that the approach has the potential to lead to significant (2-3×) gains in inference throughput.
Agentic Information Retrieval
provides an introduction to agentic information retrieval, which is shaped by the capabilities of LLM agents; discusses different types of cutting-edge applications of agentic information retrieval and challenges.
Aya Expanse
a family of open-weight foundation models for multilingual capabilities; releases an 8B and 32B parameter model, including one of the largest multilingual dataset collections to date, with 513 million examples; the release also includes Aya-101 which the authors claim is the most comprehensive multilingual models covering 101 languages; Aya Expanse 32B outperforms Gemma 2 27B, Mistral 8x22B, and Llama 3.1 70B, a model 2x its size.
A Theoretical Understanding of CoT
finds that adding correct and incorrect reasoning paths in demonstrations improves the accuracy of intermediate steps and CoT; the proposed method, Coherent CoT, significantly improves performance on several benchmarks; in the Tracking Shuffled Objects dataset, Gemini Pro shows a 6.60% improvement (from 58.20% to 64.80%), and in Penguins in a Table, DeepSeek 67B demonstrates an increase of 6.17% (from 73.97% to 80.14%).
A Survey on Data Synthesis and Augmentation for LLMs
provides a comprehensive summary of data generation techniques in the lifecycle of LLMs; includes discussions on data preparation, pre-training, fine-tuning, instruction-tuning, preference alignment, and applications.
LongRAG
enhances RAG's understanding of long-context knowledge which includes global information and factual details; consists of a hybrid retriever, an LLM-augmented information extractor, a CoT-guided filter, and an LLM-augmented generator; these are key components that enable the RAG system to mine global long-context information and effectively identify factual details; LongRAG outperforms long-context LLMs (up by 6.94%), advanced RAG (up by 6.16%), and Vanilla RAG (up by 17.25%).
Evaluation Feature Steering in LLMs
evaluates featuring steering in LLMs using an experiment that artificially dials up and down various features to analyze changes in model outputs; it focused on 29 features related to social biases and study if feature steering can help mitigate social biases; among its findings, it reports that feature steering sometimes leads to off-target effects and that a neutrality feature can help decreases social biases in 9 social dimensions without negatively affecting text quality.