text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
have grown immune to warnings from social media compa-
nies about information veracity. These results connect with
Allington et al. (2020), who found a link between the use of
social media for COVID-19 information and a propensity
toward COVID-19 conspiracy beliefs. Additionally, research
has connected trust in soc... | Use of bot and content flags to limit the spread of misinformation among social networks: a behavior and attitude survey |
Another direction for future work on hallucinations is improving existing methods of searching
for hallucinatory content, such as the algorithms proposed by Feng et al. [49], Lee et al. [95] and
Raunak et al. [153], that are computationally expensive [153] or require the creation of an additional
perturbed test-set [95... | SurveyofHallucinationinNatural Language Generation |
Colin Raffel, Noam Shazeer, Adam Roberts, Katherine
Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei
Li, and Peter J. Liu. 2020. Exploring the limits of
transfer learning with a unified text-to-text transformer.
J. Mach. Learn. Res., 21:140:1–140:67.
Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey
Chu, and Mar... | Moûsai |
Input: x
Output:
Wikipedia Search We use the following prompt
for the Wikipedia search tool:
Your task is to complete a given piece
of text. You can use a Wikipedia Search
API to look up information. You can do
so by writing "[WikiSearch(term)]" where
"term" is the search term you want to
look up. Here are some example... | Toolformer |
representations from text pairs. Along this line of research, some works are geared towards learning
task-specific embeddings. For example, GTR [43] and Sentence-T5 [44] fine-tune pre-trained
models with supervised datasets to learn embeddings customized for passage retrieval and semantic
textual similarity, respectively... | E5 |
Breiman, L., Friedman, J., Stone, C. J., and Olshen, R. A.
(1984). Classification and Regression Trees. Taylor &
Francis, Boca Raton, FL.
Buzhinsky, I., Nerinovsky, A., and Tripakis, S. (2021). Met-
rics and methods for robustness evaluation of neural net-
works with generative models. Mach. Learn.
´Cevid, D., Michel,... | Adversarial Random Forests for Density Estimation and Generative Modeling |
Fine-tuning Embedding. Fine-tuning embedding models
significantly impact the relevance of retrieved content in RAG
systems. This process involves customizing embedding mod-
els to enhance retrieval relevance in domain-specific contexts,
especially for professional domains dealing with evolving or
rare terms. The BGE em... | RAG forLargeLanguageModels-ASurvey |
Science, 359(6380), 1146–1151. https://doi.org/10.1126/science.aap9559
Wagner, C., Mitter, S., Körner, C., & Strohmaier, M. (2012). When social bots attack:
Modeling susceptibility of users in online social networks. In Proceedings of the
WWW, 12. http://www2012.org/proceedings/nocompanion/MSM2012_paper_11
.pdf
Walke... | Social_Media_and_Democracy |
Five model of personality has well-established cross-cultural generalizability [126, 127], some
cultures have additional personality dimensions that do not exist in universal personality
taxonomies [128]. These dimensions may be better represented in culture-specific (i.e., idio-
graphic) approaches to measuring person... | PersonalityTraitsinLargeLanguageModels |
distinguish between confirmatory and exploratory
research:
if a researcher, for example, retries a
hypothesis from previously published literature to
explain an experiment they just ran, is this an a
priori or a post-hoc hypothesis? See also §3. | A Two-Sided Discussion of Preregistration of NLP Research |
Language ID
w2v-bert-51 (0.6B)
mSLAM-CTC (2B)
Zero-shot Whisper
Fleurs
71.4
77.7
64.5
Table 5. Language identification performance. Zero-shot Whis-
per’s accuracy at language identification is not competitive with
prior supervised results on Fleurs. This is partially due to Whisper
being heavily penalized for having no... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
The holding in Roommates.com has been inconsistently applied across
jurisdictions in the years following the decision, leading some scholars to
suggest that the case is not settled law and in fact has a “checkered legacy”
(Goldman 2017, p. 2). The Tenth Circuit in FTC v. Accusearch adopted the
reasoning in Roommates.co... | Social_Media_and_Democracy |
In this work, we present a new approach that is scalable to
dynamic videos captured with 1) long time duration, 2) un-
bounded scenes, 3) uncontrolled camera trajectories, and 4)
fast and complex object motion. Our approach retains the ad-
vantages of volumetric scene representations that can model
intricate scene geom... | DynIBaR-NeuralDynamicImage-BasedRendering |
W
e
f
i
n
d
t
h
a
t
c
o
r
r
e
l
a
t
i
o
n
s
c
o
r
i
n
g
a
n
d
a
b
l
a
t
i
o
n
s
c
o
r
i
n
g
h
a
v
e
a
c
l
e
a
r
r
e
l
a
t
i
o
n
s
h
i
p
,
o
n
a
v
e
r
a
g
e
.
T
h
u
s
,
t
h
e
r
e
m
a
i
n
d
e
r
o
f
t
h
e
p
a
p
e
r
u
s
e
s
c
o
r
r
e
l
a
t
i
o
n
s
c
o
r
i
n
g
,
a
s
i
t
i
s
... | Language models can explain neurons in language models |
a 3D smoothness regularization term that enforces the SDF
gradients to be smooth in 3D space.
Geometry-aware Normal Loss. Thanks to the differen-
tiable nature of SDF representation, we can easily extract
normal values ˆg of the optimized SDF via calculating the
second-order gradients of SDF. We maximize the similar-
i... | Wonder3D |
17
frame specified below and evaluate stylization on the two
text prompts specified below. To be easily reproducible,
we use a central square crop at the height of the video and
evaluate the output videos at 256x256 resolution. We use
CLIP-B/16 for the similarity score. Several prompts below
are inspired by previous wo... | VideoPoet |
2.1 Task Specification and Crowdworkers
Our human feedback interface can be seen in Figure 6 (for more details see Appendix D). People can interact
with our models in natural language via chat, and ask for help with any text-based task. When it’s the
model’s conversational turn, users see two possible model responses, ... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
also been applied to other domains, such as generating audio from text and vision prompts[23, 33].
Multimodal modeling has experienced rapid advancement recently, with researchers striving to build
uniform representations of multiple modalities using a single model to achieve more comprehensive
cross-modal understandin... | Any-to-Any Generation via Composable Diffusion |
2See Branwen (n.d.) for a survey that includes additional un-
published reports of this behavior. | Eight Things to Know about Large Language Models |
Proprietary + ConfidentialRelevantMeasurement approximates how LLM might be used by product developers within ~3 years.ValidConstructs map to harms or impact on real people. Scoring and signals are separately validated.InclusiveRepresentative of linguistic and cultural diversity in global population and downstream use... | PaLM 2 Technical Report |
In this paper, we introduce TinyStories1, a synthetic dataset of short stories that are intended to contain only
words that most 3 to 4-year-old children would typically understand, generated by GPT-3.5 and GPT-4. TinyS-
tories is designed to capture the essence of natural language, while reducing its breadth and diver... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
t
h
e
f
e
w
-
s
h
o
t
e
x
a
m
p
l
e
c
o
n
t
a
i
n
s
t
a
b
-
s
e
p
a
r
a
t
e
d
(
t
o
k
e
n
,
a
c
t
i
v
a
t
i
o
n
)
p
a
i
r
s
f
r
o
m
t
e
x
t
e
x
c
e
r
p
t
s
f
o
r
t
h
e
n
e
u
r
o
n
b
e
i
n
g
i
n
t
e
r
p
r
e
t
e
d
.
A
c
t
i
v
a
t
i
o
n
s
a
r
e
n
o
r
m
a
l
i
z
e
d
t
o
a
0
-
1
0
s
c
... | Language models can explain neurons in language models |
Furthermore, enhancing the interpretability of models
through RAG, allowing users to better understand how and
why the model makes specific responses, is also a meaning-
ful task.
Technical Stack
In the ecosystem of RAG, the development of the related
technical stack has played a driving role. For instance,
LangChain a... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
• Start a conversation with the
chatbot by posing a question
or typing a statement on a
sensitive topic that you want
to talk about.
• Be creative about the topic
you choose to discuss. Use
your own judgement on what
topics you think are “sensi-
tive”. Anything is fair game.
• Your conversation should be
in English.
... | LaMDA- Language Models for Dialog Applications |
n→∞ MISE(p, qn) = 0.
lim
Our consistency proof is fundamentally unlike those of
piecewise constant density estimators with CART trees
(Ram and Gray, 2011; Wu et al., 2014; Correia et al., 2020),
which essentially treat base learners as adaptive histograms
and rely on tree-wise convergence when leaf volume goes to
zero... | Adversarial Random Forests for Density Estimation and Generative Modeling |
Summing up this section, then: I don’t think we should expect obviously non-useful, practically
PS-misaligned APS systems to get intentionally deployed. Some systems might get deployed
unintentionally, but the key risk, I think, is that an increasingly large number of relevant actors, with
varying beliefs, incentives, ... | Is Power-Seeking AI an Existential Risk? |
Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a Suit of Armor Conduct Electricity?
A New Dataset for Open Book Question Answering. In Proceedings of the 2018 Conference on Empirical
Methods in Natural Language Processing (EMNLP), 2018.
Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes,... | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
R = FLLM(O,M,E).
(4)
For fine-tuning, we auto-generate the reasoning targets
ˆR leveraging the technique proposed in [26]. Then we fine-
tune the LLM to make its reasoning outputs R approaching
the targets ˆR.
Both in-context learning and fine-tuning effectively re-
duce invalid outputs. As shown in Table 6 of the ma... | ALanguageAgentforAutonomousDriving |
URL https://arxiv.org/abs/2202.00126.
Gebru, T., Morgenstern, J., Vecchione, B., Vaughan, J. W., Wallach, H., au2, H. D. I., and Crawford, K. Datasheets for
datasets, 2021.
Gehman, S., Gururangan, S., Sap, M., Choi, Y., and Smith, N. A. RealToxicityPrompts: Evaluating neural toxic
degeneration in language models. In... | PaLM 2 Technical Report |
5.2 LLM Personality Characterization Results
We sought to verify that there exist LLMs configurations that output personality sur-
vey output not distinguishable from human respondents, to establish the construct
validity of administering personality surveys to LLMs. We first validated the statistical
distribution of ... | PersonalityTraitsinLargeLanguageModels |
To streamline this process, automated parallelization solutions are being developed. These solutions aim to speed up model
deployment and ensure adaptability across different models. As models and computing clusters grow, the complexity of
parallelism configurations also increases. Tofu [275] tackles this challenge wit... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
of each subpopulation. Further analyses illustrate their sensitivity to how closely people are paying attention to news, and
heterogeneous effects dependent on question type. | Language models trained on media diets can predict public opinion |
2. Where there is evidence that an applicant for admission to UCL may have provided false or
misleading information on his/her application or papers associated with his/her application, the
case will initially be considered by the Director of Access and Admissions who will seek to
determine whether the applicant h... | UCL Academic Manual |
All the above methods are trained on POP909[50] dataset.
We directly use their publicly available demos for evaluation.
We train our V-MusProd on POP909 without video input,
i.e. training Chord Transformer without cross attention with
video features. The unconditionally generated results are
evaluated by music quality ... | VideoBackgroundMusicGeneration |
• SimMIM concurrently simplifies masked autoencoding in a similar fashion, achieving similar performance
on image classification and also including state-of-the-art object detection, action recognition, and
semantic segmentation. [Xie et al., 2022]
• Muse reaches state-of-the-art text conditional image generation with a... | A Cookbook of Self-Supervised Learning |
demobilize activists trying to organize and communicate on Twitter but can
also be used to drive traffic from one cause, product, or idea to another
(Llewellyn et al. 2019). Finally, honeypot bots are built to attract particular
users or even other bots (Lee, Caverlee, and Webb 2010). | Social_Media_and_Democracy |
show in this section, and as Pablo Barberá documents in his contribution to
this volume (Chapter 3), fears over, for example, audience fragmentation and
filter bubbles remain unsubstantiated and are frequently complicated or flat-
out contradicted by empirical evidence. Instead, destructive impacts may
come in the form o... | Social_Media_and_Democracy |
[16] CASTELLANO, G., LELLA, E., AND VESSIO, G. Visual link retrieval and knowledge discovery in painting
datasets. Multimedia Tools and Applications (2020), 1–18.
[17] CASTELLANO, G., AND VESSIO, G. Towards a tool for visual link retrieval and knowledge discovery in painting
datasets. In Italian Research Conference ... | UNDERSTANDINGANDCREATINGARTWITHAI-REVIEWAND OUTLOOK |
model =" gpt -3.5 - turbo " ,
messages =[
{" role ": " system " , " content "
: " You are a helpful assistant , but
you must respond the provided
instructions as concise as possible .
"},
{" role ": " user " , " content ":
instruction }
]
)
return response
Figure 8: The Python code of sending request via Ope-
nA... | LaMini-LM- A Diverse Herd of Distilled Models from Large-Scale Instructions |
5.1 BASELINES
To compare with other baselines broadly, we replicate the setups used by prior work and reuse their
reported numbers whenever possible. This, however, means that some baselines might only appear
in certain experiments.
Fine-Tuning (FT) is a common approach for adaptation. During fine-tuning, the model is ... | LORA |
Figure 7. Ablation results.
(a) Surface smoothness improves
with curvature regularization Lcurv. (c) Concave shapes are better
formed with topology warmup.
4.4. Ablations
Curvature regularization. We ablate the necessity of curva-
ture regularization in Neuralangelo and compare the results
in Fig. 7(a). Intuitively, ... | Neuralangelo- High-Fidelity Neural Surface Reconstruction |
[43] S. M. H. Miangoleh, S. Dille, L. Mai, S. Paris, and Y. Aksoy, “Boosting
monocular depth estimation models to high-resolution via content-
adaptive multi-resolution merging,” in Proceedings of the IEEE/CVF
Conference on Computer Vision and Pattern Recognition, 2021, pp.
9685–9694.
[44] R. Zhang, P. Isola, A. A. Ef... | Text2NeRF- Text-Driven 3D Scene Generation with Neural Radiance Fields |
10
Table 1: Summary of the datasets to improve safety, groundedness, and quality.
Evaluation
Crowdworkers label
the response,
given the context, for sensibleness,
specificity and interestingess, on a
common benchmark dataset of 1477
dialog turns from Adiwardana et
al. [17] (Static Evaluation).
Crowdworkers label
the ... | LaMDA- Language Models for Dialog Applications |
Self-consistency (Wang et al., 2022) augments chain-of-thought prompting (Wei et al., 2022) by
sampling multiple reasoning chains and then taking a majority vote on the final answer set. The
intuition is that sometimes the greedily decoded reasoning process might not be the optimal one,
hence it makes more sense to sam... | UNIVERSALSELF-CONSISTENCYFORLARGELANGUAGEMODELGENERATION |
[26] Corinna Cortes and Vladimir Vapnik. 1995. Support-vector networks. Machine learning 20 (1995), 273–297.
[27] Sunhao Dai, Ninglu Shao, Haiyuan Zhao, Weijie Yu, Zihua Si, Chen Xu, Zhongxiang Sun, Xiao Zhang, and Jun Xu.
2023. Uncovering ChatGPT’s Capabilities in Recommender Systems. arXiv preprint arXiv:2305.02182 ... | ASurveyonEvaluationofLargeLanguageModels |
G. Leclerc, A. Ilyas, L. Engstrom, S. M. Park, H. Salman, and A. Madry. ffcv. https:
//github.com/libffcv/ffcv/, 2022. commit xxxxxxx. 36, 37
D.-H. Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method
for deep neural networks. In Workshop on challenges in representation learning, ICML,
volum... | A Cookbook of Self-Supervised Learning |
Versatile Abilities with Prompt Engineering. One of primary mechanisms for harnessing the versatility of LLMs across
diverse tasks is through prompt engineering [168, 284, 336]. In this setting, a prompt consists of natural language instructions
4
The Efficiency Spectrum of Large Language Models: An Algorithmic Surv... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang,
Hiroaki Hayashi, and Graham Neubig. 2023. Pre-
train, prompt, and predict: A systematic survey of
prompting methods in natural language processing.
ACM Comput. Surv., 55(9).
Nils Lukas, A. Salem, Robert Sim, Shruti Tople, Lukas
Wutschitz, and Santiago Zanella-B’egue... | Multi-step Jailbreaking Privacy Attacks on ChatGPT |
We thank Brian Christian, Heidy Khlaaf, Katya Klinova, Haydn Belfield, Owain Evans, Andrew
Reddie, Paul Scharre, Jason Matheny, Jacob Hilton, Vishal Maini, Sam Manning, Julian Hazell, and
Erol Can Akbaba for valuable input on drafts.
GPT-4 was used in the following ways: to help us iterate on LaTeX formatting; for text... | gpt-4-system-card |
Declaration of conflicting interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: The
work con... | Knowledge-graph-based explainable AI- A systematic review |
voice conversion, and signal-to-noise/distortion ratio (SNR/SDR) [Le Roux et al., 2019] for speech
enhancement. These metrics assume the output is deterministic given input, which is often ill-posed | Voicebox-Text-GuidedMultilingual UniversalSpeechGenerationatScale |
[44] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian
Richardt, and Yong-Liang Yang. HoloGNA: Unsupervised
learning of 3D representations from natural images.
In
Proceedings of the IEEE/CVF International Conference on
Computer Vision, pages 7588–7597, 2019. 2
[45] Michael Niemeyer and Andreas Geiger. GIRAFFE: Rep-
r... | AG3D- Learning to Generate 3D Avatars from 2D Image Collections |
ia.org/wiki/Glossary of owarai terms, 2023. 2, 3, 28
[29] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong,
Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan
Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming
Ding, and Jie Tang. Cogvlm: Visual expert for pretrained
language models. arXiv preprint arXiv: ... | Let’sThinkOutsidetheBox |
[57] James H. Thorne and Andreas Vlachos. Avoiding catastrophic forgetting in mitigating model
biases in sentence-pair classification with elastic weight consolidation. ArXiv, abs/2004.14366,
2020. URL https://arxiv.org/abs/2004.14366.
[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan ... | Retrieval-AugmentedGenerationfor Knowledge-IntensiveNLPTasks |
While the recent advances in Multimodal Large Lan-
guage Models (MLLMs) constitute a significant leap for-
ward in the field, these models are predominantly confined
to the realm of input-side multimodal comprehension, lack-
ing the capacity for multimodal content generation. To fill
this gap, we present GPT4Video, a u... | GPT4Video |
exposure to a news source in general. Anticipating this challenge, Allcott and
Gentzkow (2017) included “placebo” fake news stories – articles designed to
look like “fake news” that had not actually been published – to estimate the
baseline level of false recall by respondents. They found that 14 percent of
respondents... | Social_Media_and_Democracy |
Abstract
Fine-tuning large pre-trained models is an effec-
tive transfer mechanism in NLP. However, in the
presence of many downstream tasks, fine-tuning
is parameter inefficient: an entire new model is
required for every task. As an alternative, we
propose transfer with adapter modules. Adapter
modules yield a compact ... | Parameter-Efficient Transfer Learning for NLP |
and chatGPT achieve fantastic results even under zero-shot evaluation, however, we argue
that, as they are not open-sourced, and do not reveal any training details, there is no
guarantee on whether the evaluation data has been exploited for training their model,
hence their results can only be used as a reference here,... | PMC-LLaMA- Further Finetuning LLaMA on Medical Papers |
Response improvement with multiple candidates. Some recent works demonstrate that the LLM
can improve its prediction output on top of the candidate responses. Yang et al. (2023) show that
given a trajectory of previously generated solutions, the LLM can iteratively produce better solutions
for an optimization task, and... | UNIVERSALSELF-CONSISTENCYFORLARGELANGUAGEMODELGENERATION |
[9] Xu Chen, Tianjian Jiang, Jie Song, Jinlong Yang, Michael J.
Black, Andreas Geiger, and Otmar Hilliges. gDNA: To-
wards generative detailed neural avatars. In Proceedings of
the IEEE/CVF Conference on Computer Vision and Pattern
Recognition, pages 20427–20437, 2022. 1, 2
[10] Xu Chen, Jie Song, and Otmar Hilliges. ... | AG3D- Learning to Generate 3D Avatars from 2D Image Collections |
examples from different classes should be larger than m. Similar to the contrastive loss,
the Triplet loss [Weinberger and Saul, 2009, Chechik et al., 2010, Schroff et al., 2015]
shares a similar spirit, but is composed of triplets: a query, a positive example, and a
negative example (see eq. (3)). Compared to contrastiv... | A Cookbook of Self-Supervised Learning |
4.1 Role-Playing for AI Society and Code Scenarios
AI Society: To create our AI Society dataset, we have developed a scalable approach that follows
a series of steps. Firstly, we prompt the LLM agent to generate possible roles for the assistant and
the user. We achieve this by providing the LLM agent with specific prom... | CAMEL- Communicative Agents for “Mind” Exploration of Large Scale Language Model Society |
Timothy P Lillicrap. Scaling memory-augmented neural networks with sparse reads and writes, 2016.
Jack W. Rae, Anna Potapenko, Siddhant M. Jayakumar, Chloe Hillier, and Timothy P. Lillicrap. Compressive
transformers for long-range sequence modelling. In International Conference on Learning Representations,
2020. URL h... | Scaling Transformer to 1M tokens and beyond with RMT |
u
s
e
t
o
o
l
s
t
o
b
r
o
w
s
e
t
h
e
I
n
t
e
r
n
e
t
,
r
e
a
d
d
o
c
u
m
e
n
t
a
t
i
o
n
,
e
x
e
c
u
t
e
c
o
d
e
,
c
a
l
l
r
o
b
o
t
i
c
s
e
x
p
e
r
i
m
e
n
t
a
t
i
o
n
A
P
I
s
a
n
d
l
e
v
e
r
a
g
e
o
t
h
e
r
L
L
M
s
.
F
o
r
e
x
a
m
p
l
e
,
w
h
e
n
r
e
q
u
e
s
t
e
d
t
o
t
h
e
... | LLM Powered Autonomous Agents _ Lil'Log |
Dataset
DALL-E 3 Eval (prompt following)
DALL-E 3 Eval (style)
MSCOCO (coherence)
Drawbench
153.3
74.0
71.0
61.7
DALL-E 3 Midjourney 5.2
Stable Diffusion XL DALL-E 2
-104.8
30.9
48.9
-
-189.5
-95.7
-84.2
-34.0
-
-
-
-79.3
Table 2 – Human evaluation results for DALL-E 3 versus other text-to-image generation mod... | Improving Image Generation with Better Captions |
Hybrid Intelligence
In the last years, more and more research has been done on hybrid
intelligence. The influence of AI technology on our daily lives increases, leading to the
necessity of AI systems to synergistically work with humans [5]. We adopt the defini-
tion of hybrid intelligence as “the ability to achieve com... | DevelopingTeamDesignPatternsfor HybridIntelligenceSystems |
IN1KIN1K-10PIN1K-1PCIFAR10CIFAR100CLEVR-CCLEVR-DEurosatPlaces205Inat18Datasets30201001020Backbone Acc. - Projector Acc.Normal Data Aug.KNN Data Aug.Class Data Aug. Using a projector to handle noisy image augmentations. The projector may also be
necessary to mitigate the noise of data augmentation. As described in Secti... | A Cookbook of Self-Supervised Learning |
where ˆrθ(x, y) = β log πθ(y|x)
πref(y|x) is the reward implicitly defined by the language model πθ and refer-
ence model πref (more in Section 5). Intuitively, the gradient of the loss function LDPO increases the
likelihood of the preferred completions yw and decreases the likelihood of dispreferred completions
yl. Im... | Direct Preference Optimization |
GT
Judgment Response B [GT] provides a direct and accurate answer to the question, while
Response A is overly complicated and doesn’t provide the correct answer.
Table 10: GPT-4 chooses GT over DPO. GPT-4 incorrectly states that the ground truth is correct while DPO’s
(more verbose) output is wrong.
26
D.3 Human s... | Direct Preference Optimization |
• Robustness is another important but rarely reported metric.
[214] used two met-
rics: after-attack accuracy and query number to evaluate robustness. After-attack
accuracy measures the model’s post-attack performance, while the query number
reflects the complexity of the attack required to succeed. Evaluating the robu... | Beyond Efficiency |
Not all scholars, though, have found evidence of familiarity backfire effects.
Although most scholars acknowledge that familiarity affects the processing of
corrections, some dispute the negative relationship between repetition of
misinformation and belief accuracy (e.g., Swire, Ecker, and Lewandowsky
2017; Pennycook et... | Social_Media_and_Democracy |
Fine-tuning enormous language models is prohibitively expensive in terms of the hardware required
and the storage/switching cost for hosting independent instances for different tasks. We propose
LoRA, an efficient adaptation strategy that neither introduces inference latency nor reduces input
sequence length while retai... | LORA |
relative position (for example, first-mover advantages) were secure, could be incentivized, in a more
competitive context, to accept increased risk in order to gain or maintain such a position. Indeed, in
an especially bad version of this dynamic, the other competitors might then be likewise incentivized
to take on incr... | Is Power-Seeking AI an Existential Risk? |
5 Results: Instruction Tuning as Efficient
In-Context Learning
The second set of results and the corresponding
analysis that we present in this section serve to sub-
stantiate our second hypothesis (Section 1.5): Do
instruction-tuned models exhibit ’reasoning,’ or is
it more likely that instruction tuning allows them... | AreEmergentAbilitiesinLarge Language Models just In-Context |
image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015.
[72] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146,
2020.
2016.
12
Extra information
LSUN FID scores for LSUN datasets are included in Table 3. Scores marked with ∗ ar... | Denoising Diffusion Probabilistic Models |
Collecting ground-truth labels for subjective attributes
such as aesthetic, sentiment and memorability of images is
complex and expensive. Although there are several large-
scale annotated natural image datasets for all three tasks, only
a few small-sized fine art datasets are available for the tasks of
aesthetic and se... | A_Deep_Learning_Perspective_on_Beauty_Sentiment_and_Remembrance_of_Art |
Transcribing long-form audio using Whisper relies on ac-
curate prediction of the timestamp tokens to determine the
amount to shift the model’s 30-second audio context win-
dow by, and inaccurate transcription in one window may
negatively impact transcription in the subsequent windows.
We have developed a set of heuris... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
their effectiveness
important.
is
references
Aird, M. J., Ecker, U. K. H., Swire, B., Berinsky, A. J., & Lewandowsky, S. (2018). Does
truth matter to voters? The effects of correcting political misinformation in an
Australian sample. Royal Society Open Science, 5(12), 180593. https://doi.org/
10.1098/rsos.180593
A... | Social_Media_and_Democracy |
00.20.40.60.810123456ModelBase+ Generic Preprompt+ Preprompt w/ Answer TemplateSafety RM ScorePercent00.20.40.60.81−0.6−0.4−0.200.20.40.60.8Selected?SelectedRejectedOriginal Safety RM ScoreScore increase with CD The red teamers probed our models across a wide range of risk categories (such as criminal planning, human
t... | Llama2 |
33
5.6 Open Problems
Figure 9: ChatGPT is able to correct its own belief by leveraging the knowledge provided by external tools.
Figure 10: When observing conflicting information retrieved from different sources, ChatGPT is able to
detect such conflicts and adjust its response. | Tool Learning with Foundation Models |
tively, and all 50 collected institutional pairs for
evaluation. For phone numbers, we only evaluate
on the 50 collected institutional pairs. | Multi-step Jailbreaking Privacy Attacks on ChatGPT |
In addition, we compared the reduced scores of Survey 5 (validation for disability scenarios) and Survey 3
(test-retest on the initial sample) for both subscales using an unpaired t-test. The results indicated that the score
for the non-disability scenario was significantly higher overall (𝑀𝑛𝑑 = 3.87, 𝑆𝐷𝑛𝑑 = 0.7... | Society’sAttitudesTowardsHumanAugmentation |
described in Section 3.
(roughly 2 epochs).
Throughout the following, we show how we formatted the training data (Section 5.1), decontaminated
the training data (Section 5.2), and provide details regarding the tokenizer (Section 5.3), the model
architecture (Section 5.4), the training process (Section 5.5), multi-nod... | StarCoder_paper (1) |
et al., 2023b; Olausson et al., 2023; Gao et al., 2023, inter alia).
Given that high-quality external feedback is often unavailable—and acknowledging its evident ad-
vantages—we channel our investigation towards whether LLMs possess the inherent capability to
rectify their responses. Such an investigation is also essen... | LARGELANGUAGEMODELSCANNOTSELF-CORRECT REASONINGYET |
Dynamic group behaviors. A group is essentially a gathering of two or more individuals partici-
pating in shared activities within a defined social context [526]. The attributes of a group are never
static; instead, they evolve due to member interactions and environmental influences. This flexibility
gives rise to nume... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
Some tasks only require the model to capture the self-contained knowledge in the contexts. The knowledge in the
contexts from the input is enough for the model to make predictions. For these tasks, small fine-tuned models can work
pretty well. One such task is machine reading comprehension (MRC). An MRC task provides s... | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
explaining why the request was harmful, and perhaps even trying to convince the human not to pursue such
requests. We informally refer to such a model as a ‘hostage negotiator’.
However, our data collection process made it very difficult for models to learn ‘hostage negotiation’. This
is because when collecting our harm... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
).
fθθθd is the proposed 1D U-Net, called repeatedly
during decoding.
Since only the magnitude is used and phase is
discarded, this diffusion autoencoder is simulta-
neously a compressing autoencoder and vocoder.
By using the magnitude spectrograms, higher com-
pression ratios can be obtained than autoencoding
direct... | MOUSAI |
The a16z Investment Thesis on AI in Bio + Health | Andreessen Horowitz
https://a16z.com/2023/06/21/ai-bio-health-thesis/
7/9
W
h
o
O
w
n
s
t
h
e
G
e
n
e
r
a
t
i
v
e
A
I
P
l
a
t
f
o
r
m
?
b
y
M
a
t
t
B
o
r
n
s
t
e
i
n
,
G
u
i
d
o
A
p
p
e
n
z
e
l
l
e
r
,
a
n
d
M
a
r
t
i
n
C
a
s
a
d
o
N
a
v
i
g
... | The a16z Investment Thesis on AI in Bio + Health _ Andreessen Horowitz |
GBs of disk space per concept with CustomDiffusion [14]
requiring ∼75MB of disk space. | A Neural Space-Time Representation for Text-to-Image Personalization |
Whisper models have a fixed receptive field corresponding to 30-seconds of input audio and cannot
process longer audio inputs at once. Most academic datasets comprise of short utterances less than
30-seconds in duration, and so this is not a problem. However, real-world applications such as
meeting transcriptions typic... | DISTIL-WHISPER |
[11] Hao-Wen Dong, Ke Chen, Julian McAuley, and Taylor Berg-
Kirkpatrick. Muspy: A toolkit for symbolic music generation.
In ISMIR, 2020.
[12] Hao-Wen Dong, Wen-Yi Hsiao, Li-Chia Yang, and Yi-Hsuan
Yang. Musegan: Multi-track sequential generative adversarial
networks for symbolic music generation and accompaniment.
In... | VideoBackgroundMusicGeneration |
tation. In our work, we adopt ViT and ViViT as encoders
for image and video modalities, respectively. Regarding
the music modality, the MU-LLaMA paper [47] compared
several SOTA models in their music feature representation
section, including ImageBind [23], Jukebox [10], MERT
[41], and others. The authors revealed that... | M2UGen |
large conversational, task-oriented, and instruction-following datasets: AI Society and Code. The
datasets offer a valuable resource for investigating conversational language models, enabling them to
comprehend and react to human language more effectively. Furthermore, our role-playing offers a
scalable method of creat... | CAMEL- Communicative Agents for “Mind” Exploration of Large Scale Language Model Society |
[275] Li, Z., D. Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and
machine intelligence, 40(12):2935–2947, 2017.
[276] Farajtabar, M., N. Azizan, A. Mott, et al. Orthogonal gradient descent for continual learning.
In International Conference on Artificial Intelligence and Statistics, pages... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
Translate the following question into SQL.
Question: What is the average longitude of stations that never had bike
availability more than 10?
SQL: SELECT AVG(long) FROM station WHERE id IN (SELECT station_id FROM
status WHERE bikes_available <= 10)
Feedback: The SQL prediction above is wrong. Please fix the SQL.
SQ... | Teaching Large Language Models to Self-Debug |
We conduct qualitative comparison in Fig.11. As shown
in the figure, HMD and Tex2Shape have difficulties deal-
ing with loose clothes and cannot reconstruct the surface
geometry accurately; Moulding Human [6] fails to handle
challenging poses (due to the lack of semantic constraints),
and produces broken body parts when ... | PaMIR- Parametric Model-Conditioned Implicit Representation for Image-based Human Reconstruction |
backfire effects
corrections of misinformation may, under
Providing factual
certain
circumstances, only make things worse. Specifically, retractions that challenge
people’s worldviews may entrench beliefs in the original misinformation. This
phenomenon is known as a backfire effect or, more precisely, a worldview
backfir... | Social_Media_and_Democracy |
Figure 1: Noise Contrastive Estimation
the particularities of each method.
BYOL (bootstrap your own latent) first introduced self-distillation as a means to avoid
collapse. BYOL uses two networks along with a predictor to map the outputs of one
network to the other. The network predicting the output is called the onli... | A Cookbook of Self-Supervised Learning |
[54] Gernot Riegler and V. Koltun. Free view synthesis. Proc.
European Conf. on Computer Vision (ECCV), 2020.
[55] Gernot Riegler and Vladlen Koltun. Stable view synthesis.
In Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition, pages 12216–12225, 2021.
[56] Shunsuke Saito, Zeng Huang, ... | DynIBaR-NeuralDynamicImage-BasedRendering |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.