AI Papers of the Week
Every paper worth reading in AI, hand-picked one week at a time.
Contextual Hallucinations Mitigation in LLMs
proposes a new method that detects and significantly reduces contextual hallucinations in LLMs (e.g., reduces by 10% in the XSum summarization task); builds a hallucination detection model based on input features given by the ratio of attention weights on the context vs. newly generated tokens (for each attention head); the hypothesis is that contextual hallucinations are related to the extent to which an LLM attends to the provided contextual information; they also propose a decoding strategy based on their detection method which mitigates the contextual hallucination; the detector can also be transferred across models without the need for retraining.
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 See, LLM Do
closely investigates the effects and effectiveness of synthetic data and how it shapes a model’s internal biases, calibration, attributes, and preferences; finds that LLMs are sensitive towards certain attributes even when the synthetic data prompts appear neutral; demonstrates that it’s possible to steer the generation profiles of models towards desirable attributes.
Extracting Concepts from GPT-4
proposes a new scalable method based on sparse autoencoders to extract around 16 million interpretable patterns from GPT-4; the method demonstrates predictable scaling and is more efficient than previous techniques.
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.
Towards Scalable Automated Alignment of LLMs
provides an overview of methods used for alignment of LLMs; explores the 4 following directions: 1) aligning through inductive bias, 2) aligning through behavior imitation, 3) aligning through model feedback, and 4) aligning through environment feedback.
Extracting Interpretable Features from Claude 3 Sonnet
presents an effective method to extract millions of abstract features from an LLM that represent specific concepts; these concepts could represent people, places, programming abstractions, emotion, and more; reports that some of the discovered features are directly related to the safety aspects of the model; finds features directly related to security vulnerabilities and backdoors in code, bias, deception, sycophancy; and dangerous/criminal content, and more; these features are also used to intuititively steer the model’s output.
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.
Multimodal LLM Hallucinations
provides an overview of the recent advances in identifying, evaluating, and mitigating hallucination in multimodal LLMs; it also provides an overview of causes, evaluation benchmarks, metrics, and other strategies to deal with challenges related to detecting hallucinations.
Self-Evolution of LLMs
This survey organizes the emerging literature on self-evolving LLMs - models that improve through their own generated experience rather than additional human supervision. The authors propose a unified four-phase cycle and taxonomize existing methods across both standalone models and agent systems.
Llama 3
Meta's Llama 3 launches with 8B and 70B pretrained and instruction-tuned variants. Llama 3 8B beats Gemma 7B and Mistral 7B Instruct, and Llama 3 70B is competitive with Gemini Pro 1.5 and Claude 3 Sonnet on standard benchmarks.
Reducing Hallucination in Structured Outputs via RAG
This paper deploys a compact RAG pipeline - small retriever plus small LM - for an enterprise workflow-generation task and shows it reduces hallucination while improving out-of-domain generalization vs a baseline LLM.
Best Practices and Lessons on Synthetic Data
Google DeepMind's survey-style position paper on synthetic data for LLMs. It covers applications, quality-assurance principles, and the open challenges of factuality, fidelity, bias, and privacy.
Overview of Multilingual LLMs
A first-of-its-kind survey on multilingual LLMs, organized by multilingual alignment principles rather than model-family hierarchy. The authors propose a unified taxonomy and collect open resources to accelerate future research.
Aligning LLMs to Quote from Pre-Training Data (Quote-Tuning)
Quote-Tuning aligns LLMs to quote verbatim from trusted pre-training sources, turning the attribution step from post-hoc fact-checking into a built-in model behavior.
Many-shot Jailbreaking
Anthropic shows that long-context windows enable a new attack where hundreds of fake user/assistant dialogues are packed into a single prompt, coaxing frontier LLMs to answer the final harmful question despite safety training.
Advancing LLM Reasoning (Eurus)
OpenBMB's Eurus is a suite of reasoning-specialized LLMs (7B and 70B) fine-tuned on UltraInteract, a new alignment dataset built around preference trees for complex math, code, and logical tasks.
Logits of API-Protected LLMs Leak Proprietary Information
The paper shows that the softmax bottleneck in modern LLMs means even logit-level APIs leak enough information to reconstruct hidden architectural details.
SIMA
DeepMind's Scalable Instructable Multiworld Agent (SIMA) is a generalist AI agent that follows natural-language instructions across nine commercial 3D video games like No Man's Sky, Teardown, Valheim, and Space Engineers.
Stealing Part of a Production Language Model
The paper demonstrates the first practical attack that extracts the embedding-projection layer of production LLMs through their ordinary logit APIs.
Is Cosine-Similarity Really About Similarity?
This paper argues that cosine similarity between learned embeddings does not always measure semantic similarity, and gives analytical examples where it produces arbitrary or non-unique values.
Sora Overview
A comprehensive academic review of OpenAI's Sora, tracing the technical ingredients behind the text-to-video "world simulator" and the opportunities/limitations for the next wave of large vision models.
On the Societal Impact of Open Foundation Models
Stanford CRFM's policy paper proposes a rigorous framework for assessing the *marginal* risk of open-weight foundation models relative to closed models and pre-existing technologies.
Gemma
Google DeepMind releases Gemma, a family of open models (2B and 7B) built from the same research stack as Gemini and shipped with both base and instruction-tuned variants.