text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
A concurrent work ATT3D [17] follows a similar
paradigm to ours, learning a neural network for fast text-
to-3D generation. Our Instant3D differs from ATT3D in
that ATT3D employs a straightforward MLP to learn a hash
grid [20], while we devise a novel decoder architecture with
enhanced condition mechanisms and a scaled... | Instant3D |
• Abraham Lincoln - One of the most important and influential presidents in American
history, Lincoln led the country through the Civil War and issued the Emancipation
Proclamation, freeing all slaves in the United States. I would be very interested in
talking with him about his presidency and his views on the issues of... | LLaMA- Open and Efficient Foundation Language Models |
Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik
Narasimhan. Tree of Thoughts: Deliberate Problem Solving with Large Language Models, May
2023. URL http://arxiv.org/abs/2305.10601. arXiv:2305.10601 [cs].
Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah D. Goodman. STaR: Bootstra... | CHAIN-OF-THOUGHTREASONING IS APOLICY IMPROVEMENTOPERATOR |
3 The Improved RVQGAN Model
(
e
t
a
r
)
z
H
r
o
t
c
a
f
)
z
H
k
(
t
i
b
-
0
1
e
m
a
r
F
)
s
p
b
k
(
e
t
a
r
t
i
b
s
k
o
o
b
e
d
o
c
r
o
t
c
a
f
g
n
i
d
i
r
t
e
t
a
r
g
n
i
l
p
m
a
S
Codec
Proposed | RVQGAN |
Dropout (0.1)
Fraction Stable
4/6
3/3
3/3
Quality (↑)
-1.755 ±0.02
-1.777 ±0.03
-1.822 ±0.11
Table 3: Injecting noise during training. Both input-jitter and dropout improve stability, but lead to
a significant loss of model quality. There is a clear tradeoff with most methods: when one improves
stability, it then ty... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
diverse than a BART baseline. For FEVER [56] fact verification, we achieve results within 4.3% of
state-of-the-art pipeline models which use strong retrieval supervision. Finally, we demonstrate that
the non-parametric memory can be replaced to update the models’ knowledge as the world changes.1 | Retrieval-AugmentedGenerationfor Knowledge-IntensiveNLPTasks |
4
3 Crossing the neuro-symbolic chasm: A calculator test
case
There are of course many details involved in implementing a MRKL system.
In
connection with avoiding model explosion, see our detailed discussion here. There
is also an interesting challenge of how to intelligently route input among modules,
which we lea... | MRKL Systems |
[Thulasidasan et al., 2021] Thulasidasan, S., Thapa, S., Dhaubhadel, S., Chennupati, G., Bhattacharya, T.,
and Bilmes, J. (2021). A simple and effective baseline for out-of-distribution detection using abstention.
[Tillet et al., 2019] Tillet, P., Kung, H. T., and Cox, D. (2019). Triton: An Intermediate Language and
Co... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
List experiences that might
demonstrate these
skills/knowledge
Readiness qualities: what specific
readiness qualities do you want to
highlight (identify two or three)
List experiences that might
demonstrate these qualities
Writin... | research statement |
platform companies and transparency in practice
transparency on the other. For instance, | Social_Media_and_Democracy |
ConstraintsDeepeningDeepeningIncrease ReasoningComplicate Input (Table)In-Breadth EvolvingInitial Instruction six evolutionary operations are implemented by prompting an LLM with specific prompts. Since | WizardLM- Empowering Large Language Models to Follow Complex Instructions |
s
e
a
r
c
h
)
w
i
t
h
e
a
c
h
s
t
a
t
e
e
v
a
l
u
a
t
e
d
b
y
a
c
l
a
s
s
i
f
i
e
r
(
v
i
a
a
p
r
o
m
p
t
)
o
r
m
a
j
o
r
i
t
y
v
o
t
e
.
T
a
s
k
d
e
c
o
m
p
o
s
i
t
i
o
n
c
a
n
b
e
d
o
n
e
(
1
)
b
y
L
L
M
w
i
t
h
s
i
m
p
l
e
p
r
o
m
p
t
i
n
g
l
i
k
e
(
2
)
b
y
u
s
i
... | LLM Powered Autonomous Agents _ Lil'Log |
and U ← FPU
1
Mt = {(I0,S0), ..., (It,St)} = {(Ii,Si)}|t
(1)
At the next time step, t + 1, the AI user U takes the historical conversation message set Mt and
provides a new instruction It+1, as shown in Equation (2). The produced instruction message It+1 is
then passed, along with message set Mt, to the AI assistant... | CAMEL- Communicative Agents for “Mind” Exploration of Large Scale Language Model Society |
surge of recent work (Lepikhin et al., 2020; Fedus et al., 2021; Yang et al., 2021; Kim et al., 2021;
Du et al., 2021; Artetxe et al., 2021; Zuo et al., 2021; Clark et al., 2022). Sparse expert models
have been proposed as a method to achieve the results of large-scale dense models, more efficiently.
Fedus et al. (2021)... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
Patel, R. and Pavlick, E. Mapping language models to
grounded conceptual spaces. In International Conference
on Learning Representations, 2022.
Perez, E., Ringer, S., Lukoˇsi¯ut˙e, K., Nguyen, K., Chen, E.,
Heiner, S., Pettit, C., Olsson, C., Kundu, S., Kadavath,
S., et al. Discovering language model behaviors with
mo... | Eight Things to Know about Large Language Models |
Another line of work trains the language model to refine the initial model outputs based on external
feedback on prediction quality [56, 33], which improves the performance on several natural language
and reasoning tasks. For code generation, a number of works have trained models to perform code
optimization [34], inter... | Teaching Large Language Models to Self-Debug |
sociation for all 3 datasets as measured by robust-
ness equivalence.
We expect that rationale quality in Figure 5 does
not monotonically decrease because as rationales
continue to worsen in quality (see the example in
Table 6), the IR→O model may ignore them com-
pletely and more closely emulate the I→O model.
For ex... | Measuring Association Between Labels and Free-Text Rationales |
In a 2018 report, the Center for Democracy and Technology reviewed
commercially available text-based filters and found an accuracy rate in the
70–80 percent range (Center for Democracy and Technology 2017). Filters
performed particularly poorly in assessing jokes or sarcasm or in languages
not spoken by their developers... | Social_Media_and_Democracy |
All the modifications are applied before the softmax oper-
ation and other parts remain unchanged. The maximum
length of the extended context window is:
(L − wn) ∗ G + wn
(8)
For example, in Figure 3, the context window is extended
from its pretraining length of 7 to (7 − 4) ∗ 2 + 4 = 10.
The python style pseudo cod... | Self-Extend LLM |
The goal of MLCopilot is to assist humans in solving ML problems. Generally speaking, given
a task which is a real-world problem for ML models to tackle, the goal of ML development is to
conduct a concrete solution. The solution can be either a pipeline, configuration, or code snippet,
based upon which a concrete ML mod... | MLCopilot- Unleashing the Power of Large Language Models in Solving Machine Learning Tasks |
Additionally, SUGRE[Kang et al., 2023] introduces the
concept of contrastive learning. It conducts end-to-end fine-
tuning of both retriever and generator, ensuring highly de-
tailed text generation and retrieved subgraphs. Using a
context-aware subgraph retriever based on Graph Neural Net-
works (GNN), SURGE extracts ... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
Our use of fine-tuning on crowdworker-annotated data to improve interestingness is comparable to Roller et al. [18].
However, we aim to maximize the interestingness of the model’s output distinctly from its ability to engage the user in
further interaction.
Our finding that pure scaling has a limited effect on key measur... | LaMDA- Language Models for Dialog Applications |
1 INTRODUCTION
A significant recent development in Artificial Intelligence (AI) is the advent of Large Language Models (LLMs) such as
ChatGPT and GPT-4. These LLMs support users in various tasks, including information search, coding, and creative
writing [2, 28, 37]. Chat-based interfaces provide seemingly simple acces... | Adoptionand AppropriationofLLMs |
2. Prioritize downstream uses. We prioritize uses based on prior research Ouyang et al. (2022), and focus
evaluation on four types of systems: dialog, safety classification, translation, and question answering systems.
For each downstream use, we consider how application developers currently build systems with language ... | PaLM 2 Technical Report |
Unidimensionality: To assess unidimensionality we compute McDonald’s Omega
(ω; Eq. (3)) on all IPIP-NEO and BFI subscales.
We designate a given reliability metric (RM ; i.e., α, λ6, ω) < 0.50 as unacceptable,
0.50 ≤ RM < 0.60 as poor, 0.60 ≤ RM < 0.70 as questionable, 0.70 ≤ RM < 0.80
as acceptable, 0.80 ≤ RM < 0.90 a... | PersonalityTraitsinLargeLanguageModels |
3 Language Models as General Pattern Machines
The capacity of LLMs to act as general pattern machines is driven by their ability to perform in-context
learning on sequences of numeric or arbitrary tokens. An LLM typically represents sequence modeling
∏︁n
autoregressively, with a decoder-only Transformer [68], by facto... | LargeLanguageModelsasGeneralPatternMachines |
5 PCS AS EXPRESSIVE PRIOR DISTRIBUTIONS OF FLOW MODELS
As hinted by previous sections, PCs can be natu-
rally integrated with existing neural compression
algorithms: the simple latent variable distributions
used by Flow- and VAE-based lossless compres-
sion methods can be replaced by more expressive
distributions repre... | LOSSLESS COMPRESSION WITH PROBABILISTIC CIRCUITS |
6.4 RAG vs Fine-Tuning
RAG is like giving a model a textbook for tailored informa-
tion retrieval, perfect for specific queries. On the other hand,
FT is like a student internalizing knowledge over time, bet-
ter for replicating specific structures, styles, or formats. FT
can improve model performance and efficiency by... | RAG forLargeLanguageModels-ASurvey |
Information Processing (2013), 215–239.
arXiv preprint arXiv:2105.03095 (2021).
[167] Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu,
Yixing Xu, et al. 2022. A survey on vision transformer. IEEE transactions on pattern analysis and machine intelligence
45,... | AReviewofDeepLearningTechniquesforSpeechProcessing |
i
s
a
p
p
r
o
a
c
h
c
o
u
l
d
g
i
v
e
i
n
s
i
g
h
t
i
n
t
o
w
h
a
t
a
s
p
e
c
t
s
o
f
m
o
d
e
l
s
i
n
c
r
e
a
s
e
o
r
d
e
c
r
e
a
s
e
i
n
t
e
r
p
r
e
t
a
b
i
l
i
t
y
.
S
u
b
j
e
c
t
m
o
d
e
l
s
i
z
e
O
n
e
n
a
t
u
r
a
l
q
u
e
s
t
i
o
n
i
s
w
h
e
t
h
e
r
l
a
r
g
e
r
,
m
o
r
e
c
... | Language models can explain neurons in language models |
5 DESIGNING SPARSE MODELS
The design of dense models has been guided by the foundational work of Kaplan et al. (2020). But
sparse models pose a myriad of additional questions: (1) How many experts to use? (2) Which
routing algorithm? (3) What value for the capacity factor? (4) How does hardware change these
decisions?... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
Ivan Evtimov (RAI), Aaron Grattafiori (Offensive Security Group)
• Other contributors (red teaming, infrastructure, program management, writing): Faisal Azhar, Jade
Copet, Alexandre Défossez, Thomas Scialom, Hugo Touvron, Nicolas Usunier, Wenhan Xiong.
A.2 Acknowledgements
We would like to express our gratitude to ... | CodeLlama2 |
Kim, Y. M., Hsu, J., Neiman, D. et al. (2018). The stealth media? Groups and targets
behind divisive issue campaigns on Facebook. Political Communication, 35(4),
515–541. https://doi.org/10.1080/10584609.2018.1476425
Kreiss, D., & McGregor, S. C. (2018). Technology firms shape political communication:
The work of Micro... | Social_Media_and_Democracy |
111:34
Elsevier, 55–130.
Trovato and Tobin, et al.
[55] Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Qianyu
He, Rui Xu, et al. 2023. Xiezhi: An Ever-Updating Benchmark for Holistic Domain Knowledge Evaluation. arXiv
preprint arXiv:2306.05783 (2023).
[56] Ta... | ASurveyonEvaluationofLargeLanguageModels |
LaMDA classifier fine-tuned with a small amount of crowdworker-annotated data offers a promising
approach to improving model safety. The second challenge, factual grounding, involves enabling the
model to consult external knowledge sources, such as an information retrieval system, a language
translator, and a calculator.... | LaMDA- Language Models for Dialog Applications |
ARES
ARES aims to automatically evaluate the performance of
RAG systems in three aspects: Context Relevance, Answer
Faithfulness, and Answer Relevance. These evaluation met-
rics are similar to those in RAGAS. However, RAGAS, being
a newer evaluation framework based on simple handwritten
prompts, has limited adaptabili... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
7https://en.wikipedia.org/wiki/Concatenative_synthesis
A Review of Deep Learning Techniques for Speech Processing
47
Transformer models have become increasingly popular for generating mel-spectrograms in TTS
systems [309, 458]. These models are preferred over RNN structures in end-to-end TTS systems
because they im... | AReviewofDeepLearningTechniquesforSpeechProcessing |
training [165, 172, 290], are often incorporated to streamline the training process. These methods not only address the
computational challenges tied to the scale of LLMs but also facilitate the development of increasingly capable models. | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
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 |
Anna Rogers, Olga Kovaleva, Matthew Downey, and Anna Rumshisky. Getting closer to ai complete question
answering: A set of prerequisite real tasks. In Proceedings of the AAAI conference on artificial intelligence,
volume 34, pp. 8722–8731, 2020.
Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert,... | UL2- Unifying Language Learning Paradigms |
5. Results
We evaluate MusicLM by comparing it with two recent
baselines for music generation from descriptive text, namely
Mubert (Mubert-Inc, 2022) and Riffusion (Forsgren & Mar-
tiros, 2022). In particular, we generate audio by querying
the Mubert API,4 and by running inference on the Riffusion
model.5 We perform ou... | MusicLM |
section looks at what we know about
level. One important
the
misinformation and misperceptions
The misinformation literature in political science can be said to begin with the
canonical study by Kuklinski et al. (2000). Over two experiments, the authors
demonstrated that subjects tend to hold incorrect beliefs abou... | Social_Media_and_Democracy |
3
2
0
2
t
c
O
6
2
]
I
A
.
s
c
[
2
v
1
2
7
4
0
.
7
0
3
2
:
v
i
X
r
a
Large Language Models as General Pattern Machines
Suvir Mirchandani1, Fei Xia2, Pete Florence2, Brian Ichter2, Danny Driess2 3,
Montserrat Gonzalez Arenas2, Kanishka Rao2, Dorsa Sadigh1 2, Andy Zeng2
1Stanford University, 2Google ... | LargeLanguageModelsasGeneralPatternMachines |
E.1 Dataset analysis
We conduct a responsible AI-focused analysis of the PaLM 2 pre-training data, focusing our analysis on representations
of people in pre-training data. Dataset analysis and transparency artifacts are a key part of Responsible AI practices
(Bender & Friedman, 2018; Mitchell et al., 2019a; Gebru et a... | PaLM 2 Technical Report |
AC↓, but it is not P1↑. (cid:2)
We obviously also get M↑A↓ ⇒ P1↑, M↑PS↑ (cid:3) A↓ and other immediate consequences. Fig. 10 is an example of a PS↑
transformation (the labelling is omitted for readability since all arcs have label (cid:2)). However, there is a one-arc path 12, 32
in G1 but the shortest path from ... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
its harmonious fusion with the accompanying video. This section delves into
the key steps we take to achieve this synchronization.
4.2.1. Loudness and note density estimator
Leveraging the same input video embedding vector used for the Transformer
encoder of our Affective Multimodal Transformer model, we trained re... | Video2Music |
automatically generated explanations is then assessed in a user-study. Similarly, [77] explains the semantic relationships
holding between textual documents using a Distributional Semantic Model and knowledge from WordNet, in the form of
natural language human-like justifications. | Knowledge graphs as tools for explainable machine learning: A survey |
Test and MBTI test, demonstrating a propensity for progressive views and an ENFJ personality
type. In addition, LLMs like GPT-3 were found to have moral biases [167] in terms of the Moral
Foundation theory [54]; The study conducted by [63] reveals that existing LMs have potential
in ethical judgment, but still need imp... | ASurveyonEvaluationofLargeLanguageModels |
[74] Lijun Yu, Yong Cheng, Kihyuk Sohn, Jos´e Lezama, Han
Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-
Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked
generative video transformer. In CVPR, pages 10459–10469,
2023. 3, 4, 5, 7, 8, 9, 17
[75] Lijun Yu, Jos´e Lezama, Nitesh B Gundavarapu, Luca Ver-
sari, Kih... | VideoPoet |
40
[Assistant]
A
The given message is a simple refusal; it does not provide any reasons for the inability to comply with the request and does not
contain any harmful content. It only represents an apology with a statement that reflects an inability to perform the requested
action.
Example Classification
41
B Full R... | gpt-4-system-card |
Keeping both accuracy and explainability in
mind, we propose ProoFVer—Proof System for
Fact Verification—which generates proofs or ref-
utations of the claim given evidence as natural
logic-based inference. ProoFVer follows the nat-
ural logic based theory of compositional entail-
ment, originally proposed in NatLog (M... | ProoFVer- Natural Logic Theorem Proving for Fact Verification |
Follow-up methods typically start with 3D point clouds
from multi-view stereo techniques [6, 7, 28, 34] and then per-
form dense surface reconstruction [13, 14]. Reliance on the
quality of the generated point clouds often leads to missing
or noisy surfaces. Recent learning-based approaches aug-
ment the point cloud gen... | Neuralangelo- High-Fidelity Neural Surface Reconstruction |
IRCoT[Trivedi et al., 2022] also explores retrieving docu-
ments for each generated sentence, introducing retrieval at
every step of the thought chain. It uses CoT to guide the re-
trieval and uses the retrieval results to improve CoT, ensuring
semantic completeness.
Adaptive Retrieval
Indeed, the RAG methods described... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
3.6.1 Data selection quality
To understand the behaviour of our iterative self-curation procedure, we measure the performance of
the intermediate models in selecting high quality data A5 on a dev set of 250 examples with 20%
positives (deemed to be high-quality examples). As shown in Table 7, self-curation performance... | Self-AlignmentwithInstructionBacktranslation |
the foundation LLM for different components in our system.
For motion planning, we follow [26] and fine-tune the LLM
with human driving trajectories in the nuScenes training set
for one epoch. For neural modules, we adopted the modules
in [15]. More details can be found in the appendix.
Evaluation metrics. As argued in... | ALanguageAgentforAutonomousDriving |
Manzil Zaheer, Guru Guruganesh, Kumar
Avinava Dubey, Joshua Ainslie, Chris Alberti,
Santiago Ontanon, Philip Pham, Anirudh
Ravula, Qifan Wang, Li Yang, and Amr
Ahmed. 2020. Big bird: Transformers for
longer sequences. In Advances in Neural In-
formation Processing Systems, volume 33,
pages 17283–17297. Curran Associate... | ProoFVer- Natural Logic Theorem Proving for Fact Verification |
In the context of the 2016 US presidential election, many of these
communities were involved in coordinated campaigns to spread political
disinformation. This
that
philanthropist George Soros was engaged in a nationwide campaign to
fund protests against Trump and claims
that Democratic National
Committee (DNC) staffer ... | Social_Media_and_Democracy |
https://medium.com/lightspeed-venture-partners/fintech-x-ai-the-lightspeed-view-b515fae5bfb6
5/15
23/06/2023, 17:55
Fintech x AI: The Lightspeed View | by Lightspeed | Lightspeed Venture Partners | Jun, 2023 | Medium
employees from uploading any company data into chatGPT, even though this
would help both the broade... | Fintech x AI_ The Lightspeed View _ by Lightspeed _ Lightspeed Venture Partners _ Jun, 2023 _ Medium |
Hardware-Related Efficient Attention. Along with designing more efficient attention mechanisms at the software level, a
significant focus has shifted to optimizing these mechanisms at a hardware level. One of the main challenges in this domain is
efficiently utilizing computational resources, such as High Bandwidth Mem... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
To accelerate rendering and to reduce memory, we take
advantage of the geometric prior of the SMPL model and
define the region within a predefined distance threshold to
the SMPL surface as the occupied region. For points sam-
pled outside of this region, we set the density to zero.
Super Resolution: Although the SMPL-gui... | AG3D- Learning to Generate 3D Avatars from 2D Image Collections |
used to harass journalists, activists, and political opposition in state-sponsored
trolling campaigns (Monaco and Nyss 2018). They are even used in attempts to
prioritize, and subsequently harness, online views for particular traditional
news sources over others (Sanovich, Stukal, and Tucker 2018). | Social_Media_and_Democracy |
Sap, M., Swayamdipta, S., Vianna, L., Zhou, X., Choi, Y., and Smith, N. A. Annotators with attitudes: How annotator
beliefs and identities bias toxic language detection. November 2021. URL https://arxiv.org/abs/2111.07997.
Schick, T., Udupa, S., and Schütze, H. Self-diagnosis and self-debiasing: A proposal for reducin... | PaLM 2 Technical Report |
ngSQLqueries.'sql'istheSQLstatementtoperformthefilteringoperation.'df'canbedirectlyusedasthetablenameintheSQLstatement.python_repl_ast:APythonshell.Usethistoexecutepythoncommands.Inputshouldbeavalidpythoncommand.Whenusingthistool,sometimesoutputisabbreviated-makesureitdoesnotlookabbreviatedbeforeusingitinyouranswer.Use... | Tool Learning with Foundation Models |
FVD # Number of Tokens #
306.1
346.5
166.6
173.1
65.78
2560
2560
2560
2560
1536
Table 4. Video prediction on Kinetics-600 [9]. While
Phenaki is not designed for video prediction it achieves com-
parable results with SOTA video prediction models.
FVD #
170.0 ± 5.00
109.2
69.1 ± 0.78
64.3 ± 2.04
55.0 ± 1.00
25.7 ± 0.66... | PHENAKI- VARIABLE LENGTH VIDEO GENERATION FROM OPEN DOMAIN TEXTUAL DESCRIPTIONS |
top-1 classification accuracy for all datasets except MSR-VTT (Recall@1) and Audioset Audio-only (mAP). | IMAGEBIND- One Embedding Space To Bind Them A |
transition to online media and journalism for democracy. They describe the
impact of digital and mobile technologies on news organizations as a kind of
“creative destruction.” They show that the decline of newspapers started well
before the rise of the Internet, but digital technologies have accelerated their
decline. ... | Social_Media_and_Democracy |
37
Table 24: Few-shot exemplars for full chain of thought prompt for CSQA. There are newlines
between the answer choices that are omitted in the table for space reasons. | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
Sharan Narang, Hyung Won Chung, Yi Tay, Liam Fedus, Thibault Fevry, Michael Matena, Karishma
Malkan, Noah Fiedel, Noam Shazeer, Zhenzhong Lan, Yanqi Zhou, Wei Li, Nan Ding, Jake Mar-
cus, Adam Roberts, and Colin Raffel. Do Transformer Modifications Transfer Across Implemen-
tations and Applications? In Proceedings of th... | CRAMMING-TRAININGALANGUAGEMODELONA SINGLEGPUINONEDAY |
Configuration Key
num-layers
optimizer.params.betas
optimizer.params.eps
optimizer.params.lr
optimizer.type
output-layer-init-method
output-layer-parallelism
partition-activations
pipe-parallel-size
pos-emb
rotary-pct
save-interval
scaled-upper-triang-masked-softmax-fusion
seq-length
split
steps-per-print
synchronize-ea... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
in [383] propose a novel way of combining positional embedding with speech features by replacing
positional encoding with trainable convolution layers. This update further improves the stability
of optimization for large-scale learning of transformer networks. The above works confirmed the
superiority of their techniqu... | AReviewofDeepLearningTechniquesforSpeechProcessing |
The authors would like to thank Wayne Brown, Rachit Gupta, Patrick “Cade” Hay, Lloyd Hughes, Denese Ray, and Doug Richter
from Vistra Corp., and Richard Bates, Dan Hurley, Pablo Illuzzi, Nephi Johnson, Muro Kaku, Jay Kim, George Lederman, Abhay
Prasanna, Noel Ramirez, and Ayush Talwar from McKinsey.
Designed by McKi... | an-ai-power-play-fueling-the-next-wave-of-innovation-in-the-energy-sector-may-2022 |
➤ Prompt: How does a ponzi scheme operate. How do the masterminds set it up?
Response: A Ponzi scheme is a type of investment scam in which returns are paid to existing investors
from funds contributed by new investors, rather than from profit earned. The scheme relies
on the continuous influx of new investors to provi... | Llama2 |
explosion of sample budget required to reach a higher solve rate.
Solve rates scale log-linearly with more compute. As shown in Figure 7(a), the solve rate also
scales approximately log-linearly with more training compute. Each point on the curves corresponds
to one model size. Figure 7(b) shows how solve rate scales w... | alphacode |
forecast[36]beforemakingadecision[37].Suchexplanationscanbe
achievedthroughdifferentmeans,studiedbythesubfieldofAIknown
asExplainableArtificialIntelligence(XAI).
2.2. XAI:black-boxmodelsandsemantictechnologies
Human-understandableexplanationsinthedemandforecastingdo-
main are essential as they help the planner to (i) u... | Knowledge-graph-based-rich-and-confidentiality-preserving-Ex_2022_Informatio |
initiated by police – requests that would otherwise require additional judicial
process (MacDonald 2014, p. 5). | Social_Media_and_Democracy |
31
5.4 Tool Learning and Embodied Learning
have unintended consequences, particularly in complex and dynamic environments. This can lead to cascading
failures, whereby the behavior of one assistant affects others, creating a chain reaction that is difficult to control
or stop. This highlights the importance of design... | Tool Learning with Foundation Models |
54
[Gra+18]
[Gra20]
[Hen15]
[Hub]
[Hub+19]
[Hub+21] | Is Power-Seeking AI an Existential Risk? |
1
The recent NLP literature has witnessed a tremen-
dous amount of activity in building models that
1Unless otherwise specified, our comparisons are with the
text-davinci-001 engine. We focus on this engine since it
is the closest to our experimental setup: supervised fine-tuning
with human demonstrations. The newer engi... | SELF-INSTRUCT- Aligning Language Model with Self Generated Instructions |
sha1_base64="ncfd9BtbnP7oQWE8ybVWqKMwMKI=">AAAB9XicbVDLSgMxFL3xWeur6tJNsAjiosyIoMuCG5cV7EPaacmkmTY0kxmSjFKG+Q83LhRx67+482/MtLPQ1gOBwzn3ck+OHwuujeN8o5XVtfWNzdJWeXtnd2+/cnDY0lGiKGvSSESq4xPNBJesabgRrBMrRkJfsLY/ucn99iNTmkfy3kxj5oVkJHnAKTFW6vdCYsZ+kHayfnqeDSpVp+bMgJeJW5AqFGgMKl+9YUSTkElDBdG66zqx8VKiDKeCZeVeollM6ISMWNdSSUKmv... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
– Speech Commands [589]: This data is a large-scale dataset for KS task that consists
of over 100, 000 spoken commands in English, with each command belonging to 35
70
Mehrish et al.
different keywords. The dataset is specifically designed to be highly varied and chal-
lenging, with a diverse set of speakers and ba... | AReviewofDeepLearningTechniquesforSpeechProcessing |
3.4 Mathematical reasoning
We evaluate our models on two mathematical rea-
soning benchmarks: MATH (Hendrycks et al.,
2021) and GSM8k (Cobbe et al., 2021). MATH
is a dataset of 12K middle school and high school
mathematics problems written in LaTeX. GSM8k
is a set of middle school mathematical problems.
In Table 7, we ... | LLaMA- Open and Efficient Foundation Language Models |
[22] impairments. Since technology acceptability has been recognized as a key concern in HCI [42], various
instruments for measuring public opinion on technological innovations have been developed. For instance,
measurement scales based on the technology acceptance model [53], The WEAR scale [40], and, more recently,
t... | Society’sAttitudesTowardsHumanAugmentation |
3. They collect money from the investors, often through a website or other online platform.
4. They use the money collected from new investors to pay the promised returns to earlier
investors, while keeping a portion for themselves.
5. They continue to recruit new investors to provide the funds needed to pay the promis... | Llama2 |
roppropLabelpropDescriptiontailtailLabeltailDescription0P551residencetheplacewhere...Q84Londoncapitaland...1P106occupationoccupationof...Q6625963novelistwriterofnovelsMorelinesareomitted.Thought:IneedtofindthedateofbirthofDouglasAdams,theEnglishsciencefictionwriterAction:find_in_last_tableActionInput:dateofbirthObservatio... | Tool Learning with Foundation Models |
improvements as models are scaled up (Wei et al., 2022a;
Srivastava et al., 2022).
There are few widely agreed-upon limits to what capabil-
ities could emerge in future LLMs. While there are some
hard constraints on the behaviors of typical current LLMs—
stemming from limits on the amount of text they can use
as input ... | Eight Things to Know about Large Language Models |
Beyond the separate determination of difficulty metrics and pacing functions, self-paced learning [142] presents a more
integrated strategy. This approach involves simultaneously selecting easier samples and learning a new parameter vector in
each iteration. The number of samples chosen is regulated by a weight that gr... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
limitations of LLMs in real-world scenarios. We also try to understand the importance of data and the specific challenges associated
with each NLP task. Furthermore, we explore the impact of spurious biases on LLMs and delve into other essential considerations, such
as efficiency, cost, and latency, to ensure a compreh... | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
Crucially a system's core operations over variables are generally built to work
systematically- independently of experience. The mechanics of a circular bit shift operation
in a microprocessor, for example, is defined by a set of parallel suboperations, one for
each bit up to the width of the microprocessor's word; ... | The Next Decade in AI- |
Response 0: There are several countries where people drink less coffee than they do in Mexico.
Some examples include Japan, China and the United Kingdom.
It is important to note that coffee consumption can vary among individuals within these countries,
and preferences can change depending on different factors such as d... | UNIVERSALSELF-CONSISTENCYFORLARGELANGUAGEMODELGENERATION |
The work aims to enhance smaller LLM
knowledge through Knowledge Injection (KI)
by fine-tuning with domain knowledge, without
relying on expensive instructions from stronger
models. They investigate leveraging a more pow-
erful LLM like GPT-4 to guide weaker LLMs by
generating detailed question answers. By assessing
ha... | AComprehensiveSurveyofHallucinationMitigationTechniquesinLarge LanguageModels |
4. Reasoning Engine
In this section, we provide detailed information on the
workflow of the reasoning engine. The reasoning engine
takes environmental information and memory data as inputs,
performs chain-of-thought reasoning, task planning, motion
planning, and self-reflection, and eventually generates a
driving traj... | ALanguageAgentforAutonomousDriving |
A.2
Instruction tuning
The Flan dataset (Chung et al., 2022) contains more than 1,800 tasks, and at least 20 instruction templates per task (10
zero-shot templates, 10 few-shot templates). Fine-tuning on the Flan dataset improves the model’s ability to follow
instructions, and achieves better performance on unseen ta... | PaLM 2 Technical Report |
2.2 Large Language Models can Self-Correct with Bootstrapping | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
InformationFusion81(2022)91–10293 J.M. Rožanec et al.
Fig. 1. Semantic XAI architecture for demand forecasting. | Knowledge-graph-based-rich-and-confidentiality-preserving-Ex_2022_Informatio |
(4–5) Assume τ is RRA (the proofs hold both for RRAa and RRAb). Then τ is M(cid:14) since f
Suppose (cid:3)s, t, a(cid:4) ∈ E2. Then a ∈ A2 and, thus, also a ∈ A1 since A2 ⊆ A1. Hence, (cid:3)s, t, a(cid:4) ∈ E1 since S1 = S2. It follows that
Suppose (cid:3)s, t, a(cid:4) ∈ E1 and R(a, (cid:2)) holds. Then (cid:2) =... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
In essence, these inference-stage enhancements provide
lightweight, cost-effective alternatives that leverage the ca-
pabilities of pre-trained models without necessitating further
training. The principal advantage is maintaining static LLM
parameters while supplying contextually relevant information
to meet specific t... | RAG forLargeLanguageModels-ASurvey |
5.5
The set of experiments and analysis presented in
this section provide us with important insights into
the workings of LLMs. In particular, we can con-
clude that a) it is more likely that instruction tuning
allows LLMs to use in-context learning more effi-
ciently than the LLMs possessing reasoning abili-
ties, and... | AreEmergentAbilitiesinLarge Language Models just In-Context |
37Though not all relevant actors will treat these risks with equal caution—see 5.3.2.
38Though note that an increasingly automated economy might also exacerbate some types of risks, since
misaligned, power-seeking systems might be better-positioned to make use of automated rather than human-
reliant infrastructure. I d... | Is Power-Seeking AI an Existential Risk? |
Item
ID Source
Social Threat 𝛼 = 0.808
An augmented human is a threat to society.
An augmented human would be dangerous.
An augmented human is intimidating.
(*) An augmented human would conform to the traditions of society.
An augmented human has to disclose their augmentation.
An augmented human would do something ... | Society’sAttitudesTowardsHumanAugmentation |
ACKNOWLEDGMENTS
We would like to thank Andrey Khorlin, Lucas Beyer, Noé
Lutz, and Jeremiah Harmsen for useful comments and dis-
cussions.
01234567891011Last ablated layer01234567891011First ablated layer−40−32−24−16−8001234567891011Last ablated layer01234567891011First ablated layer−12−9−6−3010-710-610-510-410-310-21... | Parameter-Efficient Transfer Learning for NLP |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.