AI Papers of the Week
Every paper worth reading in AI, hand-picked one week at a time.
Self-Evaluation as a Defense Against Adversarial Attacks on LLMs
proposes the use of self-evaluation to defend against adversarial attacks; uses a pre-trained LLM to build defense which is more effective than fine-tuned models, dedicated safety LLMs, and enterprise moderation APIs; they evaluate different settings like attacks on the generator only and generator + evaluator combined; it shows that building a dedicated evaluator can significantly reduce the success rate of attacks.
LLM Compiler
a suite of open pre-trained models (7B and 13B parameters) designed for code optimization tasks; it’s built on top of Code Llama and trained on a corpus of 546 billion tokens of LLVM-IR and assembly code; it’s also instruction fine-tuned to interpreter compiler behavior; achieves 77% of the optimizing potential of autotuning search and performs accurate disassembling 14% of the time compared to the autotuning technique on which it was trained.
Improving Retrieval in LLMs through Synthetic Data
proposes a fine-tuning approach to improve the accuracy of retrieving information in LLMs while maintaining reasoning capabilities over long-context inputs; the fine-tuning dataset comprises numerical dictionary key-value retrieval tasks (350 samples); finds that this approach mitigates the "lost-in-the-middle" phenomenon and improves performance on both information retrieval and long-context reasoning.
Faster LLM Inference with Dynamic Draft Trees
presents a context-aware dynamic draft tree to increase the speed of inference; the previous speculative sampling method used a static draft tree for sampling which only depended on position but lacked context awareness; achieves speedup ratios ranging from 3.05x-4.26x, which is 20%-40% faster than previous work; these speedup ratios occur because the new method significantly increases the number of accepted draft tokens.
Following Length Constraints in Instructions
presents an approach for how to deal with length bias and train instruction following language models that better follow length constraint instructions; fine-tunes a model using DPO with a length instruction augmented dataset and shows less length constraint violations and while keeping a high response quality.
Adam-mini
a new optimizer that reduces memory footprint (45%-50% less memory footprint) by using fewer learning rates and achieves on-par or even outperforms AdamW; it carefully partitions parameters into blocks and assigns a single high-quality learning that outperforms Adam; achieves consistent results on language models sized from 125M -7B for pre-training, SFT, and RLHF.
TextGrad
a new framework for automatic differentiation through backpropagation on textual feedback provided by an LLM; this improves individual components and the natural language helps to optimize the computation graph; it works by providing an objective function without tuning prompts or components; claims to achieve LeetCodeHard best scores and SoTA performance on GPQA when combined with GPT4o.
Self-Tuning with LLMs
improves an LLM’s ability to effectively acquire new knowledge from raw documents through self-teaching; the three steps involved are 1) a self-teaching component that augments documents with a set of knowledge-intensive tasks focusing on memorization, comprehension, and self-reflection, 2) uses the deployed model to acquire knowledge from new documents while reviewing its QA skills, and 3) the model is configured to continually learn using only the new documents which helps with thorough acquisition of new knowledge.
Consistent Middle Enhancement in LLMs
proposes an approach to tune an LLM to effectively utilize information from the middle part of the context; it first proposes a training-efficient method to extend LLMs to longer context lengths (e.g., 4K -> 256K); it uses a truncated Gaussian to encourage sampling from the middle part of the context during fine-tuning; the approach helps to alleviate the so-called "Lost-in-the-Middle" problem in long-context LLMs.
The Geometry of Concepts in LLMs
studies the geometry of categorical concepts and how the hierarchical relations between them are encoded in LLMs; finds that simple categorical concepts are represented as simplices by the LLMs and complex concepts are represented as polytopes constructed from direct sums of simplices, which reflect the hierarchical structure.
Aligning LLMs with Demonstrated Feedback
proposes a method to align LLMs to a specific setting via a very small number of demonstrations as feedback; it aligns LLM outputs to a user’s demonstrated behaviors and can learn fine-grained style and task alignment across domains; outperforms few-shot prompting, SFT, and self-play methods on the tested benchmarks.
Financial Statement Analysis with LLMs
claims that LLMs can generate useful insights from its analysis of trends and financial ratios; shows that GPT-4 performs on par with narrowly specialized models; and achieves a profitable trading strategy based on GPT’s predictions.
Risks and Opportunities of Open-Source Generative AI
analyzes the risks and opportunities of open-source generative AI models; argues that the overall benefits of open-source generative AI outweigh its risks.
How Far Are We From AGI
presents an opinion paper addressing important questions to understand the proximity to artificial general intelligence (AGI); it provides a summary of strategies necessary to achieve AGI which includes a detailed survey, discussion, and original perspectives.
Scientific Applications of LLMs
presents INDUS, a comprehensive suite of LLMs for Earth science, biology, physics, planetary sciences, and more; includes an encoder model, embedding model, and small distilled models.
DeepSeek-Prover
introduces an approach to generate Lean 4 proof data from high-school and undergraduate-level mathematical competition problems; it uses the synthetic data, comprising of 8 million formal statements and proofs, to fine-tune a DeepSeekMath 7B model; achieves whole-proof generation accuracies of 46.3% with 64 samples and 52% cumulatively on the Lean 4 miniF2F test; this surpasses the baseline GPT-4 (23.0%) with 64 samples and a tree search RL method (41.0%).
Fine-tuning and Hallucinations
studies the impact of fine-tuning on new knowledge on the hallucination tendencies of LLMs; the setup includes fine-tuning examples that include new knowledge; shows that LLMs struggle to acquire new factual knowledge via fine-tuning; also finds that as new knowledge is learned it increases the model’s tendency to hallucinate.
WavCraft
leverages LLMs to connect task-specific models for audio content creation and editing; decomposes users' instructions into several tasks and tackles each task collaboratively with the particular module; it can enable users to interact and produce audio content without explicit commands
RLHF Workflow
provides an easily reproducible recipe for online iterative RLHF; discusses theoretical insights and algorithmic principles of online iterative RLHF and practical implementation.
AlphaFold 3
releases a new state-of-the-art model for accurately predicting the structure and interactions of molecules; it can generate the 3D structures of proteins, DNA, RNA, and smaller molecules; the model is an improved version of the Evoformer module and then assembling its predictions using a diffusion network; the diffusion process starts with a cloud of atoms which converges to its final molecular structure.
AlphaMath Almost Zero
enhances LLMs with Monte Carlo Tree Search (MCTS) to improve mathematical reasoning capabilities; the MCTS framework extends the LLM to achieve a more effective balance between exploration and exploitation; for this work, the idea is to generate high-quality math reasoning data without professional human annotations; the assumption is that a well pre-trained LLM already possesses mathematical knowledge to generate reasoning steps but needs better stimulation such as an advanced prompting or search strategy; unlike other methods such as Program-of-thought and Chain-of-thought, no solutions are required for the training data, just the math questions and the answers; the integration of LLMs, a value model, and the MCTS framework enables an effective and autonomous process of generating high-quality math reasoning data; the value model also aids the policy model in searching for effective solution paths.
Consistency LLMs
proposes efficient parallel decoders that reduce inference latency by decoding n-token sequence per inference step; the inspiration for this work comes from the human's ability to form complete sentences before articulating word by word; this process can be mimicked and learned through fine-tuning pre-trained LLMs to perform parallel decoding; it is trained to perform parallel decoding by mapping randomly initialized n-token sequences to the same result yielded by autoregressive (AR) decoding in as few steps as possible; a consistency loss helps with multiple-token prediction and a standard AR loss prevents deviation from the target LLM and ensures generation quality. Shows 2.4x to 3.4x improvements in generation speed while preserving the generation quality.
Is Flash Attention Stable?
develops an approach to understanding the effects of numeric deviation and applies it to the widely-adopted Flash Attention optimization; finds that Flash Attention sees roughly an order of magnitude more numeric deviation as compared to Baseline Attention at BF16.
MAmmoTH2
harvest 10 million naturally existing instruction data from the pre-training web corpus to enhance LLM reasoning; the approach first recalls relevant documents, extracts instruction-response pairs, and then refines the extracted pairs using open-source LLMs; MAmmoTH2-7B's (Mistral) performance increases from 11% to 34% on MATH and from 36% to 67% on GSM8K.