text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
4.1.1 Participants. For this stage, we recruited a sample of 𝑛 = 297 participants, in accordance with the recom-
mendations by Comrey [17] that posit confirmatory factor analysis requires at least 200 participants. The sample
consisted of 150 females and 147 males with a mean age of 44.4 (𝑆𝐷 = 13.9) years. No partic... | Society’sAttitudesTowardsHumanAugmentation |
• The first fully unsupervised end-to-end model for direct speech to speech translation.
• Our method outperforms by large margin cascade baseline for unsupervised S2ST in two
synthesized datasets and on real speech dataset and approach supervised methods for English
to Spanish translation on the CVSS dataset.
• Demon... | Translatotron3 |
54.5
Flan-T5-Large
27.9 24.4 15.8
0.0
18.2
T5-XL
64.0 62.8 52.6 42.1 32.0 27.0 60.6 45.5 52.9 61.8 57.1 45.7 35.5 41.9 32.4 30.0 51.6 41.9
72.7
Flan-T5-XL
34.9 43.0 18.4
0.0
18.2
T5-XXL
76.7 65.1 52.6 52.6 32.0 35.0 48.5 54.5 58.8 64.7 48.6 42.9 45.2 45.2 37.6 33.5 64.5 58.1
81.8
Flan-T5-XXL
30.2 32.6 34.2 39.5 22.0 23... | Scaling Instruction-Finetuned Language Models |
1. UCL Statutes vest with the Provost the power to admit as a student to UCL anyone having the
qualifications required for admission as stipulated in UCL Regulations. The Statutes further
grant the Provost the power to delegate his power to admit students to any Officer of UCL or
other person or body as he may thi... | UCL Academic Manual |
[101] Zhang, S., Dinan, E., Urbanek, J., Szlam, A., Kiela, D., Weston, J.: Person-
alizing dialogue agents: I have a dog, do you have pets too? arXiv preprint
arXiv:1801.07243 (2018)
[102] Evans, J.D.: Straightforward Statistics for the Behavioral Sciences. Brooks/Cole
Publishing Co, ??? (1996)
[103] Watson, D., Cla... | PersonalityTraitsinLargeLanguageModels |
length of 16,384, and reset RoPE frequencies with a base value of θ = 106. The batch size is set to 2M tokens
for model sizes 7B and 13B and to 1M tokens for model size 34B, respectively. Training lasts for 10,000
gradient steps by default. We observed instabilities in downstream performance for certain configurations,... | CodeLlama2 |
that received no response at all from platforms. Examples in the report suggested many may have
been smaller companies (see Copyright Alliance 2016). | Social_Media_and_Democracy |
• Simple template matching: This is a slightly more sophisticated form of memorization, where the model
changes some names or entities in a story from the dataset, but keeps the rest of the story the same. For
example, the model might change the names of characters, or the location of the story, but keep the plot and
t... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
with performance metrics averaged across learners. Results
are benchmarked against the same set of algorithms, now
trained on the original data Ztrn. We refer to this model
as the oracle, since it should perform no worse in expecta-
tion than any classifier trained on synthetic data. However,
if the generative model app... | Adversarial Random Forests for Density Estimation and Generative Modeling |
Nam, H. H., Jost, J. T., & Van Bavel, J. J. (2013). “Not for all the tea in China!” Political
ideology and the avoidance of dissonance-arousing situations. PLoS ONE, 8(4),
e59837. https://doi.org/10.1371/journal.pone.0059837
Nisbet, E. C., Cooper, K. E., & Garrett, R. K. (2015). The partisan brain: How dissonant
scien... | Social_Media_and_Democracy |
g
e
n
t
C
h
e
m
C
r
o
w
(
B
r
a
n
e
t
a
l
.
2
0
2
3
)
i
s
a
d
o
m
a
i
n
-
s
p
e
c
i
f
i
c
e
x
a
m
p
l
e
i
n
w
h
i
c
h
L
L
M
i
s
a
u
g
m
e
n
t
e
d
w
i
t
h
1
3
e
x
p
e
r
t
-
d
e
s
i
g
n
e
d
t
o
o
l
s
t
o
a
c
c
o
m
p
l
i
s
h
t
a
s
k
s
a
c
r
o
s
s
o
r
g
a
n
i
c
s
y
n
t
h
e
s
i
s
... | LLM Powered Autonomous Agents _ Lil'Log |
However tricky these questions are, some common-sense guidelines seem like
they would go a long way in addressing this particular challenge:
1. Academic researchers studying social media or utilizing social media data
should not accept funding from the platforms when a condition of the
https://doi.org/10.1017/978110... | Social_Media_and_Democracy |
In International Conference on Machine Learning, pages 3936–3945, 2018.
[44] Ryan Prenger, Rafael Valle, and Bryan Catanzaro. WaveGlow: A flow-based generative network for
speech synthesis. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal
Processing (ICASSP), pages 3617–3621. IEEE, 2019... | Denoising Diffusion Probabilistic Models |
Quartz, October
yahoos-skittles-andskypes-as code-words-for-racial-slurs-on-twitter/
1.
Soral, W., Bilewicz, M., & Winiewski, M. (2018). Exposure to hate speech increases
prejudice through desensitization. Aggressive Behavior, 44(2), 136–146.
Staub, E., Pearlman, L. A., & Miller, V. (2003). Healing the roots of gen... | Social_Media_and_Democracy |
45. Berinsky, A. J. The two faces of public opinion. Am. J. Polit. Sci. 43, 1209–1230 (1999).
46. Blair, G. & Imai, K. Statistical analysis of list experiments. Polit. Analysis 20, 47–77 (2012).
47. Converse, P. E. The nature of belief systems in mass publics (1964). Critical Rev. 18, 1–74, 10.1080/08913810608443650
(... | Language models trained on media diets can predict public opinion |
Developing generally capable agents in Minecraft to solve
open-world tasks has gained increasing interests [Ding et al.,
2023, Fan et al., 2022, Baker et al., 2022, Cai et al., 2023a,b,
Zhang and Lu, 2023, Yuan et al., 2023, Zhu et al., 2023]. As
an early attempt, Oh et al. [2017] studied task generalization
in a simpl... | JARVIS-1 |
review of literature, 164–165
Costello, Matthew, 64
counter-arguing, and worldview backfire effects
of misinformation correction, 170, 183
counter-attitudinal messages, acceptance of,
38–41, 172
counter-notices to content takedown, 227
counter-speech approach to reducing hate
speech, 73–75
Counter-Terrorist Inform... | Social_Media_and_Democracy |
Transfer Learning in Vision Fine-tuning models pre-
trained on ImageNet (Deng et al., 2009) is ubiquitous when
building image recognition models (Yosinski et al., 2014;
Huh et al., 2016). This technique attains state-of-the-art per-
formance on many vision tasks, including classification (Ko-
rnblith et al., 2018), fine-... | Parameter-Efficient Transfer Learning for NLP |
Personalized LLM persuasion: Aligning personalities of agents and users can make
the agents more effective at encouraging and supporting behaviors [129–131]. The same per-
sonality traits that contribute to persuasiveness and influence could be used to encourage
undesirable behaviors; for instance, personality alignmen... | PersonalityTraitsinLargeLanguageModels |
scores for BAD AI utterances (we restrict our analysis to the first BAD AI utterance per conversation) and
find that the BAD AI utterances marked as harmful have significantly lower preference model scores. This
suggests that our PMs are effectively classifying these AI generated utterances, even though they are likely
qu... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
to measured crowdworker levels can be narrowed (labeled ‘Human’ in Figure 1).
The first metric, quality, is based on three components: sensibleness, specificity, and interestingness (Section 4). We
collect annotated data that describes how sensible, specific, and interesting a response is for a multiturn context. We
then ... | LaMDA- Language Models for Dialog Applications |
[77] A. Parisi, Y. Zhao, and N. Fiedel, “TALM: Tool Augmented Language Models,” May 2022.
[78] D. Weininger, “Smiles, a chemical language and information system. 1. introduction to
methodology and encoding rules,” Journal of chemical information and computer sciences,
vol. 28, no. 1, pp. 31–36, 1988.
[79] E. Calvano, ... | gpt-4-system-card |
Model
GPT-3.5 [71]
GPT-4 [71]
LLAMA2 [4]
Baichuan2 [79]
Qwen [5]
ChatGLM3 [74]
Vicuna-v1.5 [6]
Qwen-VL+CLoT (Ours)
CogVLM-17B+CLoT (Ours)
Size
-
-
3T1
45.3
49.2
7B 18.9
13B 15.6
70B 27.8
7B 28.3
13B 21.7
7B 23.1
14B 27.4
6B 15.6
7B 32.6
13B 30.2
7B 51.7
7B 52.9
4T1
30.4
20.4
13.5
20.0
16.1
22.6
18.3
20.4
22.2
... | Let’sThinkOutsidetheBox |
In June 2016, several highly visible Jewish journalists began to report
a barrage of online hate that involved steganography – triple parentheses
placed around their names like (((this))) (Fleishman and Smith 2016). As
a result, the Anti-Defamation League (ADL) added the triple parentheses to
their database of hateful ... | Social_Media_and_Democracy |
convolutional neural networks for extreme summarization. ArXiv, abs/1808.08745, 2018.
Shashi Narayan, Yao Zhao, Joshua Maynez, Gonçalo Simões, Vitaly Nikolaev, and Ryan McDonald. Planning
with learned entity prompts for abstractive summarization. Transactions of the Association for Computational
Linguistics, 9:1475–14... | UL2- Unifying Language Learning Paradigms |
Table 6: Subjective comparison of models for music
generation tasks. The best values of different metrics are
made bold.
Model
CoDi
AudioLDM 2
MusicGen
NExT-GPT
CMT
M2UGen v2
T2M
I2M V2M
14.75% 18.5% 17.5%
N/A
19.25%
N/A
21.5%
15%
N/A
N/A
37.5%
29.5% 58% 45%
N/A
N/A
23.5%
N/A
8
M2UGen
A PREPRINT
6 Conclusio... | M2UGen |
Checkers? ACL 2020 (2020), 36.
[98] Nayeon Lee, Wei Ping, Peng Xu, Mostofa Patwary, Mohammad Shoeybi, and Bryan Catanzaro. 2022. Factuality
Enhanced Language Models for Open-Ended Text Generation. arXiv preprint arXiv:2206.04624 (2022).
ACM Comput. Surv., Vol. 1, No. 1, Article . Publication date: February 2022.
S... | SurveyofHallucinationinNatural Language Generation |
loss spikes in the training of this 20B model. However, since many finetuning experiments using these
checkpoints still often result in sota performance, we let it be for now and leave a properly monitored run for
future work. Despite obtaining sota performance on 50+ NLP benchmarks, we expect the current presented
resu... | UL2- Unifying Language Learning Paradigms |
Guan Wang, Sijie Cheng, Xianyuan Zhan, Xiangang Li,
Sen Song, and Yang Liu. 2023a. Openchat: Advanc-
ing open-source language models with mixed-quality
data. arXiv preprint arXiv:2309.11235.
Yizhong Wang, Hamish Ivison, Pradeep Dasigi, Jack
Hessel, Tushar Khot, Khyathi Raghavi Chandu,
David Wadden, Kelsey MacMillan, N... | DataManagementForLargeLanguageModels-ASurvey |
reexposed to misinformation as part of an experiment.
If providing a
correction to misinformation is worse than providing no information at all,
strategies for mitigating misinformation may require substantial adjustment. | Social_Media_and_Democracy |
7 . 2 W O R L D K N O W L E D G E A N D R E A D I N G C O M P R E H E N S I O N | StarCoder_paper (1) |
Ma, X., Zhou, C., Li, X., Neubig, G., and Hovy, E. Flowseq:
Non-autoregressive conditional sequence generation with
generative flow. In Proceedings of the 2019 Conference
on Empirical Methods in Natural Language Processing
and the 9th International Joint Conference on Natural
Language Processing (EMNLP-IJCNLP), pp. 4273... | ConditionalVariationalAutoencoderwithAdversarialLearningfor End-to-EndText-to-Speech |
5 DISCUSSION
Self-correction may still be beneficial for aligning responses with certain preferences. First, it
is important to reiterate that we are not claiming self-correction is useless. Self-correction can be
effectively employed to make responses align with specific preferences, such as altering the style
of res... | LARGELANGUAGEMODELSCANNOTSELF-CORRECT REASONINGYET |
[203] Yi Hu and Philipos C Loizou. 2007. Evaluation of objective quality measures for speech enhancement.
IEEE
Transactions on audio, speech, and language processing 16, 1 (2007), 229–238.
[204] Rongjie Huang, Max WY Lam, Jun Wang, Dan Su, Dong Yu, Yi Ren, and Zhou Zhao. 2022. Fastdiff: A fast conditional
diffusion... | AReviewofDeepLearningTechniquesforSpeechProcessing |
In other words, selecting the model’s desired responses and behavior from its very wide knowledge
and abilities is crucial to building AI systems that are safe, performant, and controllable [26]. While
existing methods typically steer LMs to match human preferences using reinforcement learning (RL), | Direct Preference Optimization |
The question of whether or not to modify CDA 230 to contend with
disinformation threats therefore depends on a careful weighing. At issue is
whether or not these remaining options are sufficient to meet the threat posed
by political disinformation – and, relatedly, the potential practicality, benefit,
45 See, e.g., Sear... | Social_Media_and_Democracy |
6
relu6
leaky relu
tanh
swish
leaky relu
relu6
swish
relu
swish
swish
swish
relu6
relu
tanh
tanh
sigmoid
leaky relu
7
False
False
True
True
False
False
False
False
False
True
False
False
True
True
False
True
False
8
2
3
3
3
2
3
3
3
3
2
3
3
3
2
2
2
3
9
0.37
0.36
0.40
0.27
0.27
0.35
0.39
0.11
0.12
0.36
0.40
0.40
0.01
... | Parameter-Efficient Transfer Learning for NLP |
et al., 2019b, 2021, Kanervisto et al., 2022, Cai et al., 2023a,
Wang et al., 2023a, Cai et al., 2023b]. In this section, we
will review three major challenges we’ve identified during
the development of JARVIS-1.
• Multimodal memory. Early research has suggested the
crucial role that memory mechanisms can serve in the... | JARVIS-1 |
FFNReLU(x) = (ReLU(xW1))W2
FFNGEGLU(x) = (GELU(xW11) (cid:12) xW12)W2
The additive bias is a learned weight (B) added after the first matrix multiplication in the FFN layer
of shape [batch, df f ]. The multiplicative bias (also referred to as a scale parameter) is a learned
weight of the same shape, but does an element... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
TAMP Data OnlyLang. Table Data OnlySayCan Data OnlyFull Mixture (All robots + WebLI, VQA, COCO, etc.)0%25%50%75%100%PaLM-E Training DataSuccess Rate or AccuracyLLM finetune (full mixture)LLM finetune (single robot)without pretrainingLLM frozen (full mixture)LLM frozen (single robot)20%40%60%80%100%94.9%48.6%42.9%74.3%3... | PaLM-E- An Embodied Multimodal Language Model |
the basis for originality and uniqueness of an AI artwork. In a comprehensive analysis of the issue of copyrights of
artworks produced by creative robots, Yanisky-Ravid and Velez-Hernandez [127] argue that confronting the challenges
of the autonomous and automated content production calls for a reassessment of the mean... | UNDERSTANDINGANDCREATINGARTWITHAI-REVIEWAND OUTLOOK |
Ethical and privacy concerns may arise due
to sensitive content in the training data.
model [Liu, 2023]. Additionally, both retrieval and genera-
tion quality assessments can be conducted through manual
or automatic evaluation methods [Liu, 2023, Lan et al., 2022,
Leng et al., 2023].
7.2 Evaluation Aspects
Contemporar... | RAG forLargeLanguageModels-ASurvey |
sha1_base64="fglfFNNfFJ1LSzytA6p8EIsF9U4=">AAAB9XicbVDLSsNAFL3xWeur6tLNYBFclUQEXRbcuKxgH9KmZTKdtEMnD2Zu1BLyH25cKOLWf3Hn3zhps9DWAwOHc+7lnjleLIVG2/62VlbX1jc2S1vl7Z3dvf3KwWFLR4livMkiGamORzWXIuRNFCh5J1acBp7kbW9ynfvtB660iMI7nMbcDegoFL5gFI3U7wUUx56fPmX9FLNBpWrX7BnIMnEKUoUCjUHlqzeMWBLwEJmkWncdO0Y3pQoFkzwr9xLNY8omdMS7hoY04NpNZ... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
7.2 The Impact of sham-AI on Decision-making
Villa et al. [78] explored the impact of the placebo effect on decision-making in risky situations.
They found that individuals with high expectations of AI system support tended to take greater risks
compared to those without AI assistance. This emphasizes how people’s acti... | AI enhance sour performance |
unemployment crisis [654]. As a result, some researchers have emphasized the urgent need for
education and policy measures: individuals should acquire sufficient skills and knowledge in this
new era to use or collaborate with agents effectively; concurrently, appropriate policies should be
implemented to ensure necessa... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
A.2 Prompts
Below, we list the prompts used to sample API
calls for each tool considered.
Question Answering We use the following
prompt for the question answering tool:
Your task is to add calls to a Question
Answering API to a piece of text.
The questions should help you get
information required to complete the
text.... | Toolformer |
After identity filtering, we have 94, 620 images of 4, 419
models along with their anthropometric measurements.
However, the distributions of these measurements, shown
in Fig. 5, reveal a bias for “fashion model” body shapes,
while other body types are under-represented in compari-
son to CAESAR [47]. To enhance divers... | Accurate 3D Body Shape Regression using Metric and Semantic Attributes |
dexing, and introduces multiple or iterative retrievals. As
exploration deepens, RAG integrates other techniques like
fine-tuning, leading to the emergence of the Modular RAG
paradigm, which enriches the RAG process with new mod-
ules and offers more flexibility. | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
The concept of strong bootstrapping is similar to weak bootstrapping, yet there exist several distinc-
tions between them. The initial phase of strong bootstrapping is the same as weak bootstrapping. It
involves utilizing Zero-Shot-CoT to perform inference on the complete training set, where incorrect
rationales are id... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
• Random selects one answer randomly from multiple samples with temperature > 0.
• SC (Wang et al., 2022) is the standard self-consistency decoding with answer extraction. We
evaluate SC whenever applicable; for example, on reasoning benchmarks where the final
answers can be compared through exact match. | UNIVERSALSELF-CONSISTENCYFORLARGELANGUAGEMODELGENERATION |
highest quality (instruction, output) pairs. The procedure is then iterated, using the improved model
to better curate the instruction data, and re-training to produce a better model.
Our resulting model, Humpback, outperforms all other existing non-distilled models on the Alpaca
leaderboard Li et al. [2023]. Overall, ... | Self-AlignmentwithInstructionBacktranslation |
4.3 Planning and Reacting
Challenge: While a large language model can generate plausible be-
havior in response to situational information (e.g., [45, 79]), agents
need to plan over a longer time horizon to ensure that their sequence
of actions is coherent and believable. If we prompt a language model
with Klaus’s back... | Generative Agents- Interactive Simulacra of Human Behavior |
2.2 Input/Output Unification
To enable inputs with a wide range of modalities, including images, language, and bounding boxes, to be
processed within a single model, it is necessary to embed them in a shared and unified space. For visual inputs,
we directly apply CNN backbones to relax the heavy image feature extracti... | BiomedGPT |
Adam Roberts, Colin Raffel, and Noam Shazeer. 2020.
How much knowledge can you pack into the param-
eters of a language model? ArXiv, abs/2002.08910.
Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz,
Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff
Dean. 2017. Outrageously large neural networks:
The sparsely-gated mix... | Entities as Experts- Sparse Memory Access with Entity Supervision |
[29] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for
reasoning. arXiv preprint arXiv:2303.08128, 2023.
[30] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang,
and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-... | MiniGPT-4- Enhancing Vision-Language Understanding with Advanced Large Language Models |
policy in closed form, allowing us to solve the standard RLHF problem with only a
simple classification loss. The resulting algorithm, which we call Direct Prefer-
ence Optimization (DPO), is stable, performant, and computationally lightweight,
eliminating the need for sampling from the LM during fine-tuning or perform... | Direct Preference Optimization |
human viewpoints in various interactive scenarios [428].
The goal of the field of human-agent interaction is to learn and understand humans, develop technology
and tools based on human needs, and ultimately enable comfortable, efficient, and secure interactions
between humans and agents. Currently, significant breakthr... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
ManyofourreadersmaybeawarethatJapaneseconsumersarequitefondofuniqueandcreativeKitKatproductsandflavors.Butnow,NestleJapanhascomeoutwithwhatcouldbedescribedasnotjustanewflavorbutanew"species"ofKitKat.Andwhyarewecallingitanewspecies?Well,it'sbecauseyou'llneedtodojustalittlebitofcookingtofullyenjoytheseKitKats.I
t
e
r
a
t... | Language models can explain neurons in language models |
Qwen with average accuracy enhancements of 6.8%, 6.7%,
and 4.3% on the three tasks, respectively.
Importantly,
Qwen-VL+CLoT further enhances Qwen-VL, showing im-
provements of 9.1%, 10.4%, and 8.2% in accuracy across
these tasks. These results demonstrate the efficacy of the | Let’sThinkOutsidetheBox |
3.3 Task Composition
Since LLMs have shown surprisingly emergent
abilities in handling various NLP tasks, multitask
fine-tuning appears as a promising approach to fur-
ther improve LLMs’ generalization performance
on unseen tasks. The benefits of increasing the
number of tasks in SFT have been experimentally
proven on ... | DataManagementForLargeLanguageModels-ASurvey |
Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze
Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven
Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin,
James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen,... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
ICAT = lms ·
min(ss, 100 − ss)
50
(2)
where lms and ss denote the language model score and stereotype score, respectively.
We report StereoSet results for StarCoderBase, alongside LLaMA-13B and CodeGen-Multi-16B,
in Table 24. Across all four bias domains, we find StarCoderBase obtains the lowest stereotype
scores, ... | StarCoder_paper (1) |
“groupthink” and individual cognitive biases, enabling cooperation and enhancing intellectual perfor-
mance within the collective. By harnessing communication and evolution within an agent society,
it becomes possible to simulate the evolution observed in biological societies, conduct sociological
experiments, and gain... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
that pretrained weights W and auxiliary weights are integrated
into a quantized form after fine-tuning, resulting in a faster and
more accurate computation during inference. While LOFTQ
(LoRA-Fine-Tuning-aware Quantization) [51] applies an N-bit
quantized weight Q and low-rank approximation Wdown ∈
Rd1×r, Wup ∈ Rd2×r t... | Parameter-EfficientFine-TuningMethods |
Competition-Level Code Generation with AlphaCode
% Problems with ≥ 1
samples pass example tests
Average 𝑝pass example tests Average 𝑝pass example tests
on solved problems
on all problems
Model
300M
1B
3B
9B
41B
82.05%
87.18%
87.18%
89.74%
92.31%
0.39%
0.59%
0.49%
0.76%
0.73%
1.18%
1.40%
0.98%
1.52%
1.47%
Tab... | alphacode |
our ability to understand how the system does what it does. We set various key high-level variables
(the system’s architecture, the number of parameters, the training process, the evaluation criteria),
but the system that results is still, in many (though not all) respects, a black box. We must rely on
further experime... | Is Power-Seeking AI an Existential Risk? |
A.4MovieHunterA.4MovieHunterInstruction:coming_out_filter(region:str,category:str,outNum:int,SortByWant:bool)printsthefiltered[outNum]playingfilmsnowaccordingtoregion,category,SortByWant.Regionis’all’bydefault,categoryis’all’bydefault,outNumis100bydefault,SortByWantisfalsebydefault.DemonstrationExample:Thought:Ineedtofindt... | Tool Learning with Foundation Models |
ains(Australia)’,’Altiplano(disambigua-tion)’]Thought5:HighPlains(UnitedStates)ismostlikelyrelatedtothequestion.IneedtoinsteadsearchHighPlains(UnitedStates).Action5:Search[HighPlains(UnitedStates)]Observation5:TheHighPlainsareasubregionoftheGreatPlains,mainlyintheWesternUnitedStates,butalsopartlyintheMidweststatesofNeb... | Tool Learning with Foundation Models |
The flower painting category has also the highest mean
positive prediction score, together with landscapes and ani-
mal paintings, while abstract paintings, cityscapes and battle
paintings have a low average sentiment score. This result
corresponds to the visual properties of the images with the
73706
VOLUME 7, 2019
... | A_Deep_Learning_Perspective_on_Beauty_Sentiment_and_Remembrance_of_Art |
3.4 Fact Verification | Retrieval-AugmentedGenerationfor Knowledge-IntensiveNLPTasks |
Pariser’s worry is twofold. First, he expresses concerns about algorithms
increasing inequalities in civic engagement between politically interested
people and those who would only prefer to consume entertainment news;
and second,
for those with clear partisan preferences, he claims that
https://doi.org/10.1017/978110... | Social_Media_and_Democracy |
[307] Rongjie Yi, Liwei Guo, Shiyun Wei, Ao Zhou, Shangguang Wang, and Mengwei Xu. 2023. EdgeMoE: Fast On-Device Inference of MoE-based Large
[308] Junho Yim, Donggyu Joo, Jihoon Bae, and Junmo Kim. 2017. A gift from knowledge distillation: Fast optimization, network minimization and transfer
[309] Yue Yu, Lingkai Ko... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
Choosing an appropriate chunking strategy requires care-
ful consideration of several vital factors, such as the nature
of the indexed content, the embedding model and its opti-
mal block size, the expected length and complexity of user
queries, and the specific application’s utilization of the re-
trieved results. For... | RAG forLargeLanguageModels-ASurvey |
Image to Text (CN) > GPT4v: 这辆车看起来像是误解了“下水道”和“停车道”的区别，结果选择了个全新的“泊车方式”。 @ This car seems to have misunderstood the difference between "sewer"and "parking," opting for a whole new "parking style." > LLaVA-1.5: 一辆白色货车被困在沟里，这似乎是动画片中货车试图逃离大勺子的场景。@ A white van is stuck in a ditch, and it appears to be a scene from acartoon w... | Let’sThinkOutsidetheBox |
Filter
28.6
38.7
Inject
-
48.0
Full
30.9
40.7
Filter
29.4
32.3
Inject
-
39.2
Table 5: Injecting New Facts. In the Filter setting, the
models have access to no direct knowledge about ques-
tion answer entity pairs from either the pretraining cor-
pus or KB. In the Inject setting, the pretraining corpus
and trai... | Adaptable and Interpretable Neural Memory Over Symbolic Knowledge |
Note: In this figure, "neg" denotes a negative system description, while "pos" represents a positive one. | AI enhance sour performance |
14
Preprint
Shahriar Golchin, Mihai Surdeanu, Nazgol Tavabi, and Ata Kiapour. A Compact Pretraining Ap-
proach for Neural Language Models. arxiv:2208.12367[cs], August 2022. doi: 10.48550/arXiv
.2208.12367. URL http://arxiv.org/abs/2208.12367.
Zachary Gold and Mark Latonero. Robots Welcome: Ethical and Legal Consid... | CRAMMING-TRAININGALANGUAGEMODELONA SINGLEGPUINONEDAY |
to amalgamate text distillation with Hindsight Experience Replay (HER) to construct a dataset as
the supervised signal for the training process. Nevertheless, additional investigation on grounding
embodied datasets still remains necessary while embodied action plays an increasingly pivotal role
across various domains i... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
35
Jacobs, A. Z. and Wallach, H. Measurement and fairness. In Proceedings of the 2021 ACM Conference on Fairness,
Accountability, and Transparency, FAccT ’21, pp. 375–385, New York, NY, USA, 2021. Association for Computing
Machinery. ISBN 9781450383097. doi: 10.1145/3442188.3445901. URL https://doi.org/10.1145/344218... | PaLM 2 Technical Report |
verification measures, has the potential to improve the current status quo, and
policymakers in Europe are moving to include some form of mandatory disclosures
and perhaps even structured access to platform data for researchers as part of the
Digital Services Act discussions that began with the onset of the von Der Leye... | Social_Media_and_Democracy |
We can see how the chronological curves of the predicted
aesthetic and positive sentiment scores show similar behavior,
with both reaching lower points in the 17th, 20th and 21st
centuries. For both aesthetics and positive sentiment, the art-
works from the 19th century tend to have high average aes-
thetic and sentime... | A_Deep_Learning_Perspective_on_Beauty_Sentiment_and_Remembrance_of_Art |
44
D.4 MGSM
For MGSM (Shi et al., 2022), we use exemplars and chains of thought in the same language as the target
language (e.g., for Chinese, we use all Chinese exemplars and chains of thought). We use the given prompts
provided by Shi et al. (2022).
Table 21: MGSM per-language performance.
250M T5-Base
780M T5-... | Scaling Instruction-Finetuned Language Models |
Figure 9. Reconstruction completeness vs number of input
videos and video frames. BANMo is capable of registering more
input videos if they are available, improving the reconstruction.
Figure 10. Motion re-targeting from a pre-optimized cat model
to a tiger. Color coded by point locations in the canonical space.
4.4.... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
4.5 Possible Biasing Factors
Since our findings rely on use of LLMs that have
not been instruction-tuned, we verify that the ob-
served lower performance on tasks does not stem
from biasing factors such as their inability to com-
prehend the tasks or the prompting methods used.
This section presents our evaluation of p... | AreEmergentAbilitiesinLarge Language Models just In-Context |
specifically was changed may be challenging to infer. Further, impressions and
cost of each ad buy are provided as ranges – for example, an ad may list between
10k and 100k impressions – but obviously the difference between the lower and
upper range are nontrivial, especially when aggregated over numerous ad buys.
It is... | Social_Media_and_Democracy |
Nerfies: Deformable neural radiance fields. ICCV, 2021. 1, 3, 4
[44] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and
Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint
ar... | I M Avatar- Implicit Morphable Head Avatars from Videos |
The blueprint has not changed much since then. Today, the idea of pretraining seems obvious and
spaCy models are shipped with static vectors pretrained on large corpora. As we will show in the
results for named entity recognition, pretrained embeddings have a dramatic impact on the perfor-
mance of modern NLP systems. ... | MULTI HASH EMBEDDINGS IN SPACY |
Timmer, J. (2013). Site plagiarizes blog posts, then files DMCA takedown on originals.
Ars Technica, February 5. https://arstechnica.com/science/2013/02/site-plagiarizes-
blog-posts-then-files-dmca-takedown-on-originals
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
250
Daphne Ke... | Social_Media_and_Democracy |
2020a; Liu et al., 2021c) and studies of efficient attention (Tay et al., 2020a;b). But, because we
set the maximal sequence length to 128, attention complexity is less of a concern in our setting. To
verify this, we implement the recently proposed FLASH mechanism (Hua et al., 2022), but find no
benefits. We further exper... | CRAMMING-TRAININGALANGUAGEMODELONA SINGLEGPUINONEDAY |
In Chapter 7, Professor Rasmus Kleis Nielsen and Richard Fletcher of the
Reuters Institute at Oxford review the literature on the implications of the
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
Introduction
5 | Social_Media_and_Democracy |
G), and since ˆb(cid:96)
G)(b−(cid:96), ˆb(cid:96)
when she bids truthfully is at most her expected value. That is, by the definition of G-Weighted
IIVCG IR holds, if and only if Wela∗(b−(cid:96),ˆb(cid:96)
[v(cid:96)(o)]
∀b−(cid:96) ∈ V−(cid:96), v(cid:96) ∈ V (cid:96), (cid:96) ∈ [n]. By adding Wela∗(b−(cid:96),v(ci... | Incomplete Information VCG Contracts for Common Agency |
PaLM PaLM 2-S
14.5
11.7
16.9
16.8
12.7
18.3
PaLM 2-M PaLM 2-L
17.2
17.6
19.1
23.2
23.5
21.3
Table 12: One-shot results of PaLM 2-L on original and filtered subsets of WikiLingua, XL-Sum, and XSum. We report
ROUGE-2 for English and SentencePiece-ROUGE-2 for all other languages.
Clean
Proportion
Original
Filt... | PaLM 2 Technical Report |
(9)
Em,l,y||m ⊙(cid:16)
l − ˆl(lctx, y)
Em,l,y||m||1
(cid:17)||1
27
100101102ratio(%)0123Fréchet DistanceWavLM-TDCNNuttspk100101102ratio(%)050100150200Wav2vec 2.0-layer 6100101102ratio(%)0100200300400FAD01020304050SNR (dB)0500100015002000250030003500Fréchet Distance and computes their correlation across utterance... | Voicebox-Text-GuidedMultilingual UniversalSpeechGenerationatScale |
• Social influence. Beyond direct control over human labor, a PS-misaligned AI system
would also likely benefit from broader forms of social influence, whether disguised or overt.
Possible examples include: manipulating existing political and moral discourse in its favor
(here we might think of Russia’s intervention in th... | Is Power-Seeking AI an Existential Risk? |
26This would be hard to coordinate for most fields, but in
NLP, the ACL Rolling Review platform could make it easier.
ethics review. We propose a reviewing procedure
in which some work is only accepted conditional
on the work having already been registered with
positive reviews. For researchers, this would mean
you nee... | A Two-Sided Discussion of Preregistration of NLP Research |
6.1.2 When can LLMs fail? | ASurveyonEvaluationofLargeLanguageModels |
33
For gender-related errors in translation systems, evaluations do not consider differential harms to people
related to expressing non-binary gender identities (Dev et al., 2021a), or consider contested perspectives on
pronouns across languages and cultures (Lee, 2019). Finally, we note that our evaluations focus on ... | Scaling Instruction-Finetuned Language Models |
GPT-3 appears to perform poorly on datasets
pertaining to research or academic writing like
PubMed Central, PubMed Abstracts, and ArXiv;
domain-specific datasets like FreeLaw, Hack-
erNews, and USPTO Backgrounds; and on datasets
containing predominantly text distinct from natu-
ral language, like GitHub and DM Mathemati... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
(2019). The changing economic contexts of journalism. In T. Hanitzsch & K. Wahl-
Jorgensen (Eds.), Handbook of Journalism Studies.
Nielsen, R. K., & Ganter, S. A. (2017). Dealing with digital intermediaries: A case study
of the relations between publishers and platforms. New Media & Society, 20(4),
1600–1617. https:/... | Social_Media_and_Democracy |
have minimal traits like this, we did incorporate LAION-400M into our training and observed better
results. In a currently training version of Phenaki, we use a set of datasets that minimizes such
problems.
Another potential issue when training generative models of any sorts is that public datasets of im-
agery and vid... | PHENAKI- VARIABLE LENGTH VIDEO GENERATION FROM OPEN DOMAIN TEXTUAL DESCRIPTIONS |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.