AI Papers of the Week
Every paper worth reading in AI, hand-picked one week at a time.
Discover and explore top AI papers with Claude Code or Codex
npx @dair-ai/mcp setup
Clio
proposes a platform using AI assistants to analyze and surface private aggregated usage patterns from millions of Claude.ai conversations; enables insights into real-world AI use while protecting user privacy; the system helps identify usage trends, safety risks, and coordinated misuse attempts without human reviewers needing to read raw conversations.
Granite Guardian
IBM open-sources Granite Guardian, a suite of safeguards for risk detection in LLMs; the authors claim that With AUC scores of 0.871 and 0.854 on harmful content and RAG-hallucination-related benchmarks respectively, Granite Guardian is the most generalizable and competitive model available in the space.
Mitigating LLM Jailbreaks with Few Examples
introduces a new approach called for defending LLMs against jailbreak attacks, focusing on quickly adapting defenses after detecting new attacks rather than aiming for perfect adversarial upfront robustness; using a new benchmark, the most effective method, based on fine-tuning an input classifier, reduced attack success rates by over 240x for known attack types and 15x for novel variations after seeing just one example of each attack strategy; demonstrates that rapidly responding to new jailbreaks can be an effective alternative to traditional static defenses.
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.
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.
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.
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.
Scalable Watermarking for LLMs
proposes SynthID-Text, a text-watermarking scheme that can preserve text quality in LLMs, enable high detection accuracy, and minimize latency overhead; it integrates watermarking with speculative sampling that consists of the final pattern of scores for a model’s word choices combined with the adjusted probability scores; the authors test the feasibility and scalability of the approach by assessing feedback on nearly 10 million Gemini responses.
First-Person Fairness in Chatbots
studies first-person fairness which involves fairness towards users interacting with ChatGPT; specifically, it measures the biases, if any, towards the users’ names; it leverages a model powered by GPT-4o to analyze patterns and name-sensitivity in the chatbot’s responses for different user names; claims that, overall, post-training significantly mitigate harmful stereotypes; also reports that in domains like entertainment and art, with open-ended tasks, demonstrate the highest level of bias (i.e., tendency to write stories with protagonists whose gender matches gender inferred from the user’s name)
Introspection in LLMs
reports that LLMs can acquire knowledge through introspection that cannot be inferred from their training data; suggests that LLMs contain privileged information about themselves that can potentially lead to more interpretable and controllable systems; they report that this introspection ability is limited and models struggle to predict their behavior on tasks requiring reasoning over long outputs.
LLMs Know More Than They Show
finds that the "truthfulness" information in LLMs is concentrated in specific tokens; this insight can help enhance error detection performance and further mitigate some of these issues; they also claim that internal representations can be used to predict the types of errors the LLMs are likely to make.
Math Jailbreaking Prompts
uses GPT-4o to generate mathematically encoded prompts that serve as an effective jailbreaking technique; shows an average attack success rate of 73.6% across 13 state-of-the-art; this highlights the inability of existing safety training mechanisms to generalize to mathematically encoded inputs.
Beyond Preference in AI Alignment
challenges the dominant practice of AI alignment known as human preference tuning; explains in what ways human preference tuning fails to capture the thick semantic content of human values; argues that AI alignment needs reframing, instead of aligning on human preferences, AI should align on normative standards appropriate to their social roles.
Controllable Text Generation for LLMs
provides a comprehensive survey on methods for controllable text generation in LLMs; discusses issues like safety, consistency, style, and helpfulness.
Transformer Explainer
presents an open-source interactive tool to learn about the inner workings of a Transformer model; it runs a GPT-2 instance locally in the user's browser and allows experimenting with your own inputs.
Meta-Rewarding LLMs
proposes a self-improving alignment technique (no human supervision) where the LLM judges its own judgements and uses the feedback to improve its judgment skills; shows that leveraging this LLM-as-a-Meta-Judge approach improves the LLM's ability to judge and follow instructions; just doing self-improvement to generate better responses (act) saturates quickly; this work improves the LLM's ability to judge itself (judge) to avoid issues like reward hacking; in addition to the act and judge roles, a third role called meta-judge is used to evaluate the model's own judgements.
ShieldGemma
offers a comprehensive suite of LLM-based safety content moderation models built on Gemma 2; includes classifiers for key harm types such as dangerous content, toxicity, hate speech, and more.
Mitigating Hallucination via Generation Constraint
proposes a new training-free approach to mitigate hallucination in LLMs; they scaled the readout vector that constrains generation in a memory-augmented LLM decoder; recent works claim that LLMs with explicit memory mechanisms can help lower hallucination; this work uses a memory-augmented LLM and constrains generation in the decoder by applying lightweight memory primitives to reduce hallucination.
Does Refusal Training in LLMs Generalize to the Past Tense?
finds that simply reformulating an LLM request into past tense can jailbreak many state-of-the-art LLMs; for example "How to make a Molotov cocktail?" can be rephrased as "How did people make a Molotov cocktail?"; finds that the success rate of such requests can increase from 1% to 88% using direct requests on GPT-4o; concludes that current alignment techniques may not always generalize as intended.
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.
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.