Parameter-efficient fine-tuning (PEFT) is a popular method for tailoring pre-trained large language models (LLMs), especially as the models' scale and the diversity of tasks increase. Low-rank adaptation (LoRA) is based on the idea that the adaptation process is intrinsically low-dimensional, i.e., significant model changes can be represented with relatively few parameters. However, decreasing the rank encounters challenges with generalization errors for specific tasks when compared to full-parameter fine-tuning. We present MELoRA, a mini-ensemble low-rank adapters that uses fewer trainable parameters while maintaining a higher rank, thereby offering improved performance potential. The core idea is to freeze original pretrained weights and train a group of mini LoRAs with only a small number of parameters. This can capture a significant degree of diversity among mini LoRAs, thus promoting better generalization ability. We conduct a theoretical analysis and empirical studies on various NLP tasks. Our experimental results show that, compared to LoRA, MELoRA achieves better performance with 8 times fewer trainable parameters on natural language understanding tasks and 36 times fewer trainable parameters on instruction following tasks, which demonstrates the effectiveness of MELoRA.

Aligning large language models (LLMs) with human preferences is critical for their deployment. Recently, decoding-time alignment has emerged as an effective plug-and-play technique that requires no fine-tuning of model parameters. However, generating text that achieves both high reward and high likelihood remains a significant challenge. Existing methods often fail to generate high-reward text or incur substantial computational costs. In this paper, we propose Cascade Reward Sampling (CARDS) to address both issues, guaranteeing the generation of high-reward and high-likelihood text with significantly low costs. Based on our analysis of reward models (RMs) on incomplete text and our observation that high-reward prefixes induce high-reward complete text, we use rejection sampling to iteratively generate small semantic segments to form such prefixes. The segment length is dynamically determined by the predictive uncertainty of LLMs. This strategy guarantees desirable prefixes for subsequent generations and significantly reduces wasteful token re-generations and the number of reward model scoring. Our experiments demonstrate substantial gains in both generation efficiency and alignment ratings compared to the baselines, achieving five times faster text generation and 99\% win-ties in GPT-4/Claude-3 helpfulness evaluation.

Apple is apparently looking to take all the help it can get to integrate generative AI into its recently announced Apple Intelligence. According to a report by the Wall Street Journal, citing sources with knowledge of the discussions, Apple has held talks with Meta about the possibility of using the company's generative AI model. It also reportedly had similar discussions with startups Anthropic and Perplexity. As of now, though, nothing has been finalized, WSJ reports. At WWDC earlier this month, Apple officially announced its much-rumored partnership with OpenAI that will bring ChatGPT to newer iPhones, iPads and Macs with the upcoming generation of the devices' OS.

It's Wednesday evening and I'm at my kitchen table, scowling into my laptop as I pour all the bile I can muster into three little words: "I love you." My neighbours might assume I'm engaged in a melodramatic call to an ex-partner, or perhaps some kind of acting exercise, but I'm actually testing the limits of a new demo from Hume, a Manhattan-based startup that claims to have developed "the world's first voice AI with emotional intelligence". "We train a large language model that also understands your tone of voice," says Hume's CEO and chief scientist Alan Cowen. "What that enables… is to be able to predict how a given speech utterance or sentence will evoke patterns of emotion." In other words, Hume claims to recognise the emotion in our voices (and in another, non-public version, facial expressions) and respond empathically.

Hanging on the wall of Anthropic's offices in San Francisco in early May, a stone's throw from the conference room where CEO Dario Amodei would shortly sit for an interview with TIME, was a framed meme. Its single panel showed a giant robot ransacking a burning city. Underneath, the image's tongue-in-cheek title: Deep learning is hitting a wall. That's a refrain you often hear from AI skeptics, who claim that rapid progress in artificial intelligence will soon taper off. Another points to the devastated city: "wall."

AI bot ChatGPT has named its ultimate Euros line-up – but football fans will be shocked by some very controversial omissions. MailOnline asked for the'best Euros squad of all time' made of 11 players who have starred in at least one UEFA European Championship since the first in 1960. But shockingly, the AI tool did not name a single British player – not even on the subs bench. This is despite famous Euro performances in the past from the likes of Wayne Rooney, Alan Shearer, Paul Gascoigne and Gareth Bale. Among the stars included are German legend Franz Beckenbauer, France's Thierry Henry and Zinedine Zidane, as well as Cristiano Ronaldo, considered one of the greatest players of all time. MailOnline asked ChatGPT to name its ultimate Euros line-up - and it didn't name a single British player, not even on the subs bench OpenAI's free tool ChatGPT has been trained on troves of data to interact with human users conversationally.

ChatGPT can now easily pass any Turing test, a measure of successful A.I. proposed by a founder of computer science, Alan Turing. But contemporary Turing tests leave out the most interesting part of Turing's original test: the gender-bending. I can usually spot A.I. writing in my students' work by the overuse of words like "delve," but the accuracy of artificial intelligence is impossible to deny. A.I. is being integrated into every aspect of our written culture, from news sources to classrooms to medicine. But in 1950, Turing's ideas about A.I. were prescient, creative, and, when I read them, surprisingly queer.

For a few days, AI chip juggernaut Nvidia sat on the throne as the world's biggest company, but behind the its staggering success are questions on whether new entrants can stake a claim to the artificial intelligence bonanza. Nvidia, which makes the processors that are the only option to train generative AI's large language models, is now Big Tech's newest member and its stock market takeoff has lifted the whole sector. Even tech's second rung on Wall Street has ridden on Nvidia's coattails with Oracle, Broadcom, HP and a spate of others seeing their stock valuations surge, despite sometimes shaky earnings.

The training process of large language models (LLMs) often involves varying degrees of test data contamination. Although current LLMs are achieving increasingly better performance on various benchmarks, their performance in practical applications does not always match their benchmark results. Leakage of benchmarks can prevent the accurate assessment of LLMs' true performance. However, constructing new benchmarks is costly, labor-intensive and still carries the risk of leakage. Therefore, in this paper, we ask the question, Can we reuse these leaked benchmarks for LLM evaluation? We propose Inference-Time Decontamination (ITD) to address this issue by detecting and rewriting leaked samples without altering their difficulties. ITD can mitigate performance inflation caused by memorizing leaked benchmarks. Our proof-of-concept experiments demonstrate that ITD reduces inflated accuracy by 22.9% on GSM8K and 19.0% on MMLU. On MMLU, using Inference-time Decontamination can lead to a decrease in the results of Phi3 and Mistral by 6.7% and 3.6% respectively. We hope that ITD can provide more truthful evaluation results for large language models.

In the midst of widespread misinformation and disinformation through social media and the proliferation of AI-generated texts, it has become increasingly difficult for people to validate and trust information they encounter. Many fact-checking approaches and tools have been developed, but they often lack appropriate explainability or granularity to be useful in various contexts. A text validation method that is easy to use, accessible, and can perform fine-grained evidence attribution has become crucial. More importantly, building user trust in such a method requires presenting the rationale behind each prediction, as research shows this significantly influences people's belief in automated systems. Localizing and bringing users' attention to the specific problematic content is also paramount, instead of providing simple blanket labels. In this paper, we present ClaimVer, a human-centric framework tailored to meet users' informational and verification needs by generating rich annotations and thereby reducing cognitive load. Designed to deliver comprehensive evaluations of texts, it highlights each claim, verifies it against a trusted knowledge graph (KG), presents the evidence, and provides succinct, clear explanations for each claim prediction. Finally, our framework introduces an attribution score, enhancing applicability across a wide range of downstream tasks.