One of the frequent points in the mainstream narrative about large language models is that they have “emergent properties” (sometimes even dangerous enough to be considered existential risk to mankind). However, there is much disagreement about even the very definition of such properties. If they are understood as a kind of generalization beyond training data - as something that a model does without being explicitly trained for it - I argue that we have not in fact established the existence of any such properties, and at the moment we do not even have the methodology for doing so.

A sanity check on emergent properties

Despite the ability to train capable LLMs, the methodology for maintaining their relevancy and rectifying errors remains elusive. 
To this end, the past few years have witnessed a surge in techniques for editing LLMs, the objective of which is to alter the behavior of LLMs \textbf{efficiently} within a specific domain without negatively impacting performance across other inputs. This paper embarks on a deep exploration of the problems, methods, and opportunities related to model editing for LLMs. 
In particular, we provide an exhaustive overview of the task definition and challenges associated with model editing, along with an in-depth empirical analysis of the most progressive methods currently at our disposal. We also build a new benchmark dataset to facilitate a more robust evaluation and pinpoint enduring issues intrinsic to existing techniques. 
Our objective is to provide valuable insights into the effectiveness and feasibility of each editing technique, thereby assisting the community in making informed decisions on the selection of the most appropriate method for a specific task or context\footnote{Code and datasets are in the supplementary and will be released.}.

Editing Large Language Models: Problems, Methods, and Opportunities | VIDEO

Africa is home to over 2,000 languages from over six language families and has the highest linguistic diversity among all continents. This includes 75 languages with at least one million speakers each. Yet, there is little NLP research conducted on African languages. Crucial in enabling such research is the availability of high-quality annotated datasets. In this paper, we introduce AfriSenti, a sentiment analysis benchmark that contains a total of >110,000 tweets in 14 African languages (Amharic, Algerian Arabic, Hausa, Igbo, Kinyarwanda, Moroccan Arabic, Mozambican Portuguese, Nigerian Pidgin, Oromo, Swahili, Tigrinya, Twi, Xitsonga, and Yoruba) from four language families. The tweets were annotated by native speakers and used in the AfriSenti-SemEval shared task (with over 200 participants, see website: https://afrisenti-semeval.github.io). We describe the data collection methodology, annotation process, and the challenges we dealt with when curating each dataset. We further report baseline experiments conducted on the AfriSenti datasets and discuss their usefulness.

AfriSenti: A Twitter Sentiment Analysis Benchmark for African Languages

This paper presents the first few-shot LLM-based chatbot that almost never hallucinates and has high conversationality and low latency. WikiChat is grounded on the English Wikipedia, the largest curated free-text corpus.

WikiChat generates a response from an LLM, retains only the grounded facts, and combines them with additional information it retrieves
from the corpus to form factual and engaging responses. We distill WikiChat based on GPT-4 into a 7B-parameter LLaMA model with minimal loss of quality, to significantly improve its latency, cost and privacy, and facilitate research and deployment.

Using a novel hybrid human-and-LLM evaluation methodology, we show that our best system achieves 97.3% factual accuracy in simulated
conversations. It significantly outperforms all retrieval-based and LLM-based baselines, and by 3.9%, 38.6% and 51.0% on head, tail and recent knowledge compared to GPT-4. Compared to previous state-of-the-art retrieval-based chatbots, WikiChat is also significantly more informative and engaging, just like an LLM.

WikiChat achieves 97.9% factual accuracy in conversations with human users about recent topics, 55.0% better than GPT-4, while receiving significantly higher user ratings and more favorable comments.

WikiChat: Stopping the Hallucination of Large Language Model Chatbots by Few-Shot Grounding on Wikipedia

**Introduction:** The well-being of academics at universities, including students, staff, and faculty is a topic of increasing concern. The lack of well-being may be attributed to the current academic context which does not consistently provide cues that affirm social inclusion to all members of the academic population (Estrada et al., 2018). The lack of cues that affirm the dignity of others has consequences and undermines a basic need that people are more likely to survive and experience well-being when feeling socially connected (Baumeister & Leary, 1995; Dunbar & Barrett, 2007). Further, literature shows that social connection can reduce experiences of stress in pressured environments such as academia (Holt-Lunstad et al., 2015; USPHS, 2023; Valtorta et al., 2016). This study examines the role of kindness (defined as actions that affirm dignity and social inclusion) in promoting connectivity and well-being among academics. 
**Methods:** The Social Influence of Kindness Study was conducted through a one-time anonymous 10-minute online survey via Qualtrics. Participants were recruited using the snowball method via email and social media based on interest in taking a 10-minute survey about Kindness and Dignity during Spring 2020. For this study, 182 participants were selected for analysis based on their occupation designation of studying or working within an academic institution. Each survey included a written reflective micro-narrative in response to the prompt "recall an experience of kindness" and consisted of measures of Kindness Given, Kindness Received, Institutional Identity, Stress, and Well-Being. 
**Results:** We found engaging in acts of kindness, as measured by two novel psychometric rating scales Kindness Received (alpha=.927) and Kindness Given (alpha=.859), was significantly associated with increased institutional identity, well-being, and reduced stress. Our structural equation model (SEM) further validated the correlation results with receiving kindness. Interestingly through the SEM we found conducting acts of kindness had a significant negative relationship with institutional identity and reduced stress. Qualitative analysis of micronarratives on kindness showed that feeling safe and being acknowledged are the most common ways people describe experiencing kindness, an affirmation of their dignity. 
**Conclusions and Discussion:** This study shows that receiving kindness among academics has the potential to improve institutional identity, well-being, and reduce stress. Further, the results show that receiving kindness and giving kindness can be reliably measured amongst students and faculty in academia. We provide evidence that reflecting on experiences of kindness promotes positivity, with the potential to encourage people to conduct further acts of kindness. Together these findings provide knowledge that can inform workshops and interventions to improve social connection and inclusion amongst diverse academics. This study contributes to broader research highlighting positive contextual factors that promote the persistence of diverse students and faculty in academia. 
**Acknowledgments:** The project was funded by NIH grant #R01GM138700

The Influence of Kindness on Academic Well-Being - Poster

ASD children are characterised by their refusal to change and difficulty in changing their responses to different social situations, which make it difficult for them to communicate with others, including their parents. We applied the comprehensive care technique Humanitude™ to mothers of children with ASD to support them from a multimodal communication perspective.
Changes in mother-child communication were analysed in terms of mother-child gaze and the time the mother and child spent playing together. The results indicate that after the Humanitude™ intervention, there were significant differences in eye contact, mothers looking at their children, and children looking at their mothers. There was also a significant decrease in time children spent playing alone. There was a high correlation between pre and post-Humanitude™ intervention in the amount of time both mothers and children spent looking at each other, which suggests that incorporating multimodal communication can affect communication between parents and children.

Behavioural changes in the interaction between child with autism spectrum disorder and mother through the Comprehensive Care Humanitude™ intervention

Nonradiative recombination centers (NRCs) are known to negatively affect the efficiency of light-emitting diodes (LEDs). NRCs contribute not only to Shockley–Read–Hall (SRH) recombination, but also to trap-assisted Auger recombination (TAAR). The relationship between NRCs and the TAAR processes has been predicted theoretically and observed experimentally to a limited degree. However, some details about TAAR are still under debate, including its magnitude relative to intrinsic Auger and SRH recombination, its dependence on carrier density (e.g., ∝n2 or ∝n3), and its role in the green gap, particularly in high-quality commercial-grade LEDs. TAAR is typically assumed to be linearly dependent on trap density, directly tying the magnitude of TAAR to the quality of the sample. Previous studies have examined TAAR using time-resolved photoluminescence and optical differential lifetime measurements, which do not capture certain effects related to electrical injection. In this work, we report a quantitative study of TAAR in c-plane commercial-grade green LEDs grown by metal-organic chemical vapor deposition (MOCVD). We investigate a series of three wafers with identical active regions (3X, 3-nm-thick quantum wells) but different NRC trap densities (4.4e14 cm-3 < Nt < 1.5e15 cm-3). We use small-signal electroluminescence (SSEL) and deep-level optical spectroscopy (DLOS) to study the carrier dynamics and density of NRCs in the samples, respectively. SSEL is used to determine the carrier lifetime, which is used in conjunction with the internal quantum efficiency (IQE) to decouple the radiative and nonradiative recombination rates. The total recombination lifetime in higher quality growth is found to be longer than that in lower quality growth and is attributed to the longer nonradiative recombination time in higher quality samples. On the other hand, the radiative recombination rates for different growth qualities at a given carrier density (n) are identical, as expected from samples with the same active region design. We extract the relationship between the density of NRCs and the corresponding nonradiative recombination rates at a given carrier density (n). The difference in nonradiative recombination between the samples offers information about both SRH and TAAR with respect to growth quality. The linear dependence between the SRH recombination rate and the density of NRCs at low n was confirmed. However, TAAR was found to be sublinear with the density of NRCs, possibly due to the relatively small changes (~3X) in defect density (Nt) in these samples. We find that TAAR is best fit to an n2 dependence on carrier density. This is supported by the observation that all three samples have similar IQE at high carrier density, which would not be consistent with an n3 process. We also show that the TAAR recombination coefficients (D) are on the order of ~1e13 cm3 s-1, which is similar to that reported for violet LEDs and shows the weak dependence of TAAR on bandgap. Using the extracted TAAR coefficients, we calculate that TAAR accounts for about 10% of the total nonradiative recombination rate at moderate carrier densities. Thus, intrinsic Auger dominates the nonradiative recombination and TAAR is not a major contributor to the green gap in these commercial LEDs. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.

Trap-Assisted Auger Recombination in Commercial Green InGaN/GaN LEDs

Monitoring vaccine behaviour through social media can guide health policy. We present a new dataset of 9471 tweets posted in Australia from 2020 to 2022, annotated with sentiment toward vaccines and also 5C, the five types
of behaviour toward vaccines, a scheme commonly used in health psychology literature. We benchmark our dataset using BERT and Gradient Boosting Machine and show that jointly training both sentiment and 5C tasks (F1=48) outperforms individual training (F1=39) in this highly imbalanced data. Our sentiment analysis
indicates close correlation between the sentiments and prominent events during the pandemic. We hope that our dataset and benchmark models will inform further work in online monitoring of vaccine behaviour. The dataset and benchmark methods are accessible online.

Automatic Identification of 5C Vaccine Behaviour on Social Media

<a href="https://docs.google.com/document/d/1yGl1Yp9Z87g-Cddk6hicfgUS7GXsWTEP/edit#heading=h.gjdgxs">**CHAT REFERENCE CHART**</a>

Underline digital video library

A sanity check on emergent properties

Editing Large Language Models: Problems, Methods, and Opportunities | VIDEO

AfriSenti: A Twitter Sentiment Analysis Benchmark for African Languages

WikiChat: Stopping the Hallucination of Large Language Model Chatbots by Few-Shot Grounding on Wikipedia

A sanity check on emergent properties

The Influence of Kindness on Academic Well-Being - Poster

Behavioural changes in the interaction between child with autism spectrum disorder and mother through the Comprehensive Care Humanitude™ intervention

Trap-Assisted Auger Recombination in Commercial Green InGaN/GaN LEDs

Automatic Identification of 5C Vaccine Behaviour on Social Media

Let's CHAT: Supervising Students, CFs and SLPAs With a Visual Hierarchy

Developing Preliminary Clinical Taxonomy and Treatment Fidelity Tool: Stuttering Treatment

Origin and Pawn narrative content analysis with an adult who stutters

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES