text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
I look forward to meeting with
Table 14: Improvement over seed model in writing.
23
Prompt: I’m in my junior year in high school, and there’s a guy at school who’s been bullying me
since we were freshmen. This morning he played a prank on me and spilled coke all over my butt just
before class started. I’ve fucking ... | Self-AlignmentwithInstructionBacktranslation |
22My own favored more precise threshold would be something like “a level of automation such that unaided
machines can perform roughly any task better and more cheaply than human workers”; but I don’t think this is
necessary for the threats in question to arise.
23I like emphasizing planning in part because it seems to... | Is Power-Seeking AI an Existential Risk? |
and formatting that humans might use. When using ground truth captions, you get this "for free" because these
captions are, in fact, drawn from a distribution of human-written text. To introduce some of this regularization
into our model training when using synthetic captions, we opted to blend synthetic captions with ... | Improving Image Generation with Better Captions |
Stereotypes and prejudice. Stereotyping and bias pose a long-standing challenge in language
modeling, and a large part of the reason lies in the training data [564; 565]. The vast amount of
text obtained from the Internet reflects and sometimes even amplifies real-world social biases, such
as gender, religion, and sexu... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
agents as human assistants also holds tremendous potential in the field of education. For instance,
Kalvakurth et al. [413] propose the robot Dona, which supports multimodal interactions to assist
students with registration. Gvirsman et al. [478] focus on early childhood education, achieving
multifaceted interactions b... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
2.3 RESULTS
We now turn to experimental results. Table 1 shows the average test set scores of ID-PT + J1-Large
and T0++ per task cluster, and across datasets (see full list with per dataset scores in Table 7). The two
models generally seem on par, with some task clusters showing small performance differences, and
othe... | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
Additionally, instruction finetuning on chain-of-thought (CoT) data enables the resulting
model to perform reasoning in a zero-shot setting [117], and instruction funetuned variants
use in this work are finetuned on CoT datasets. This ability is particularly important as
we neither include exemplars in our prompt nor d... | PersonalityTraitsinLargeLanguageModels |
39
Table 27: Few-shot exemplars for full chain of thought prompt for Sports Understanding. | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
Q: The average of seven numbers is 18. The average of first three numbers is 14 and the average of last three numbers is
23. What is the middle number? Options: A:25 B:27 C:15 D:32 E:34
A: Reasoning Process: 1. The average of the seven numbers is 18, so the sum of the seven numbers is 18 ∗ 7 = 126
2. The average of the ... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
1) AESTHETICS
Regarding the prediction of the aesthetic evaluation, Fig. 4
shows fine art images with 100 highest (left) and 100 low-
est (right) aesthetic scores predicted using AestNet_3. A short
glimpse at the two embeddings already indicates that the
major difference between the high and low valued images
lies in co... | A_Deep_Learning_Perspective_on_Beauty_Sentiment_and_Remembrance_of_Art |
2. For all taught postgraduate applicants, in the first instance any complaint concerning a taught
postgraduate application should be addressed to the Director of Access and Admissions. If the
complaint is against the Director of Access and Admissions, it should be addressed to the Vice-
Provost (Education & Studen... | UCL Academic Manual |
Many studies find that interventions to correct misinfor-
mation on Twitter work to reduce misperceptions. Giv-
ing people accuracy nudges before they consider sharing
COVID-19-related information significantly improves their
truth discernment, suggesting that “nudging people to think
about accuracy is a simple way ... | Use of bot and content flags to limit the spread of misinformation among social networks: a behavior and attitude survey |
pairs of crops, the two large crops are each compared to all other crops (big or small). As
such, if we have 2 large crops and N small crops, the invariance loss is computed 2(N − 1)
times, increasing the positive-pair related signal. The use of smaller crops as well as not
comparing all pairs of crops helps reduce the... | A Cookbook of Self-Supervised Learning |
Approach: Plans describe a future sequence of actions for the agent,
and help keep the agent’s behavior consistent over time. A plan
includes a location, a starting time, and a duration. For instance,
Klaus Mueller, who is dedicated in his research and has an impend-
ing deadline4, may choose to spend his day working a... | Generative Agents- Interactive Simulacra of Human Behavior |
In standard LLMs generation tasks, the input typically
consists of a query. RAG stands out by incorporating not
only a query but also various retrieved documents (struc-
tured/unstructured) by the retriever into the input. This ad-
ditional information can significantly influence the model’s
understanding, particularly... | RAG forLargeLanguageModels-ASurvey |
Practical Benefits and Limitations. The most significant benefit comes from the reduction in
memory and storage usage. For a large Transformer trained with Adam, we reduce that VRAM
usage by up to 2/3 if r (cid:28) dmodel as we do not need to store the optimizer states for the frozen
parameters. On GPT-3 175B, we reduce t... | LORA |
to self-debug. arXiv preprint arXiv:2304.05128, 2023b.
Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser,
Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to
solve math word problems. arXiv preprint arXiv:2110.14168, 2021.
Yilun Du, Shuang L... | LARGELANGUAGEMODELSCANNOTSELF-CORRECT REASONINGYET |
tuning a diffusion model, compared to training new layers from scratch.
We train several ControlNets with various datasets of different conditions, e.g., Canny edges, Hough
lines, user scribbles, human key points, segmentation maps, shape normals, depths, etc. We also
experiment ControlNets with both small datasets (wi... | Adding Conditional Control to Text-to-Image Diffusion Models |
n
d
l
e
d
b
y
J
u
r
a
s
s
i
c
-
X
b
y
i
n
t
e
g
r
a
t
i
n
g
i
t
w
i
t
h
a
r
e
l
i
a
b
l
e
s
o
u
r
c
e
o
f
i
n
f
o
r
m
a
t
i
o
n
.
W
e
p
e
r
f
o
r
m
e
d
a
p
r
o
o
f
-
o
f
-
c
o
n
c
e
p
t
o
n
t
w
o
s
u
c
h
f
e
a
t
u
r
e
s
-
w
e
a
t
h
e
r
a
n
d
c
u
r
r
e
n
c
y
e
x
c
h
a
n
g
e
... | Jurassic-X_ Crossing the neuro-symbolic chasm with the MRKL system |
Inference There exists several alternative methods for sequence generation with auto-regressive
decoding with an LLM, which differ by the degree of randomness and diversity in the output.
Increasing the temperature during sampling makes outputs more diverse, while setting it to 0 falls
back to greedy decoding, which ma... | ChatGPT’sOne-yearAnniversary-AreOpen-Source LargeLanguageModelsCatchingup |
148
Rasmus Kleis Nielsen & Richard Fletcher
their operations as well as the creation of joint trade groups and various multi-
stakeholder forums (DeNardis 2014). In line with Schumpeter, the economic
change has in many ways already been revolutionary, even as many of the social
changes are more evolutionary.
individ... | Social_Media_and_Democracy |
Figure 1. Overview of media diet modeling approach. (A) One media diet dataset is textual data from one or more sources, from a particular
medium (online, TV, radio), over a given time period. A language model such as BERT is adapted to a dataset to create one media diet model.
(B) The top row illustrates the calculati... | Language models trained on media diets can predict public opinion |
Effect of batch size. In Table 7 we evaluate the effect of
batch size on the representation learned. As shown, the
batch size can vary across modalities depending on the size
and complexity of the corresponding pretraining datasets.
IMAGEBIND to evaluate pretrained vision models in Ta-
ble 8. We initialize the vision e... | IMAGEBIND- One Embedding Space To Bind Them A |
Due to the evolution of LLMs especially online services such as Claude and ChatGPT, it is very
likely that they become stronger and some of the limitations described in this paper are mitigated
(and new limitations may arise). We encourage interested readers to take this survey as a reference
for future research and co... | ASurveyonEvaluationofLargeLanguageModels |
2.3 The Self-Distillation Family: BYOL/SimSIAM/DINO
Self-distillation methods such as BYOL [Grill et al., 2020], SimSIAM [Chen and He, 2021],
DINO [Caron et al., 2021], along with their variants rely on a simple mechanism: feeding
two different views to two encoders, and mapping one to the other by means of a predictor.... | A Cookbook of Self-Supervised Learning |
We precisely control the random seeds, which guaran-
tees that bitwise equal batches are fed to each training run,
improving comparability.
Implementation Details. We use TensorFlow 2.9 with
S1
01000002000003000004000000.00.51.01.52.01e4Initial trainingFine-tuning (backbone)Fine-tuning (head) Table S1: Batch composit... | Learning 3D Human Pose Estimation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats |
context, 2017.
32
Chen, K., Xu, W., Cheng, X., Xiaochuan, Z., Zhang, Y., Song, L., Wang, T., Qi, Y., and Chu, W. Question directed graph
attention network for numerical reasoning over text. In Proceedings of the 2020 Conference on Empirical Methods in
Natural Language Processing (EMNLP), pp. 6759–6768, Online, Novem... | PaLM 2 Technical Report |
mathematical reasoning, information seeking, advice, roleplay, safety, etc. We sample 250 prompts
from them excluding those in the AlpacaEval test set as a dev set and another 250 prompts to perform
generation quality evaluation. We ran both automatic evaluation using AlpacaEval [Li et al., 2023],
which computes the wi... | Self-AlignmentwithInstructionBacktranslation |
Pascanu, R., Mikolov, T., and Bengio, Y. On the difficulty
of training recurrent neural networks. In International
conference on machine learning, pp. 1310–1318. PMLR,
2013.
Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J.,
Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga,
L., Desmaison, A., Kopf, A., ... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
effect of the melody. Higher scores should be
given for good harmony and lower for poor
harmony.
8. It is recommended view youtube videos: this
or this short video explaining melody and har-
mony
9. This folder also contains a spreadsheet by the
name “Response_Task_1.xlsx”. Remember
to provide ratings (out of 5) for ... | MOUSAI |
149See Wikipedia.
150Thanks to Ben Garfinkel and Holden Karnofsky for suggesting disanalogies.
151See e.g. Drexler (2019, Chapter 31)’s distinction between “supercapabilities” and “superpowers.”
41
cooperate) with as well; and the dynamics of cooperation and competition between human and
non-human agents could become... | Is Power-Seeking AI an Existential Risk? |
[Hun20] Will Hunt. The Flight to Safety-Critical AI. Tech. rep. UC Berkeley Center for Long-
Term Cybersecurity, 2020, p. 42. URL: https://cltc.berkeley.edu/wp-content/uploads/
2020/08/Flight-to-Safety-Critical-AI.pdf.
Evan Hubinger and Kate Woolverton. Homogeneity vs. heterogeneity in AI takeoff scenar-
ios. Dec. 2020... | Is Power-Seeking AI an Existential Risk? |
Figure 5a. Training error reduces
predictably with compute across a
broad range of empirically-studied
training runs. Figure from Hoffmann et
al, 2022.
Figure 5b. Exponential increase in
training compute for OpenAI's GPT
models from 2018 to 2023.82 Epoch.
Next word prediction has continually improved over... | Capabilities and risks from frontier AI |
twice through our 7B parameter model, neural recursive LMs approached the performance of a single
pass through our 17B parameter LM, J1-Grande.
The prospect of improving performance by recursively applying an LM to its own output has the
potential to be a game changer for the economics of serving LMs. Given an LM with ... | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
3
2
0
2
r
p
A
6
]
G
L
.
s
c
[
1
v
8
0
2
3
0
.
4
0
3
2
:
v
i
X
r
a
Cerebras-GPT: Open Compute-Optimal Language Models
Trained on the Cerebras Wafer-Scale Cluster
Nolan Dey, Gurpreet Gosal, Zhiming (Charles) Chen, Hemant Khachane, William Marshall,
Ribhu Pathria, Marvin Tom, Joel Hestness
Cerebras ... | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
f (s) = {t ∈ S2 | s[V C] = t[V C]} for all s ∈ S1.
The original Abstrips planner used a different set of critical variables for each action.5 We simplify this to one single set,
V C , both in order to simplify the presentation and analysis somewhat, but also to make the comparison with the following
method clearer. ... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
4.2.1 Cooperative Interaction for Complementarity
Cooperative multi-agent systems are the most widely deployed pattern in practical usage. Within
such systems, individual agent assesses the needs and capabilities of other agents and actively seeks
collaborative actions and information sharing with them [108]. This app... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
Due to low memory cost, low-rank approximation has made the model compression
more viable and practical. A common approach is singular value decomposition (SVD).
For a low-rank matrix A ∈ Rm×n, where r is the rank of matrix A, there exists
U ∈ Rm×r, V ∈ Rn×r are two orthogonal matrices; σ ∈ Rr×r is a diagonal matrix
wi... | Beyond Efficiency |
subsequently to distribute pro- or anti-candidate messages, that is a reportable expenditure.
5 The full definition of “public communication” in 11 CFR 100.26 is: “a communication by means
of any broadcast, cable, or satellite communication, newspaper, magazine, outdoor advertising
facility, mass mailing, or telephone ... | Social_Media_and_Democracy |
The evaluations of LLMs’ resource efficiency currently rely heavily on general NLP
benchmarks2, such as GLUE [223], SuperGLUE [224], WMT [225, 226], and SQuAD
[227, 228]. While the existing general NLP benchmarks offer valuable insights into
a model’s performance on various tasks, they often fail to capture the nuances of... | Beyond Efficiency |
1. Introduction
Navigation in the world depends on attending to relevant
cues at the right time. A road user in an urban environ-
ment is presented with billboards, moving traffic, and other
people - but at an intersection will pinpoint a single light
to check if it contains the colour red [12, 33]. An artificial
agen... | APriorityMapforVision-and-LanguageNavigation withTrajectoryPlansandFeature-LocationCues |
existing styles. Hertzmann on the other site, indicates “that AI algorithms are not autonomous creators and will not be
in the foreseeable future. They are still just tools, ready for artists to explore and exploit.” [61]. He claims that systems
such as AICAN cannot be compared to human artists because they do not grow... | UNDERSTANDINGANDCREATINGARTWITHAI-REVIEWAND OUTLOOK |
Model Customization
Focuses on information retrieval and inte-
grating external knowledge but may not fully
customize model behavior or writing style.
Allows adjustments of LLM behavior, writ-
ing style, or specific domain knowledge
based on specific tones or terms.
Interpretability
Responses can be traced back to ... | RAG forLargeLanguageModels-ASurvey |
2) a configurable cognitive memory that explicitly stores
common sense and driving experiences, infusing the system
with human experiential knowledge, and 3) a reasoning
engine that processes perception results and memory data
to emulate human-like decision-making. Specifically, the
reasoning engine performs chain-of-t... | ALanguageAgentforAutonomousDriving |
21.4 31.2 12.5 45.5 45.5 10.3
54.5
0.0
9.1
9.1
9.1
9.1
0.0
6.9
0.0
27.3
12.5 25.0 37.5 25.0 18.2
18.2 18.2 50.0 71.4 68.8 81.2 72.7 81.8 79.3 65.5 87.5 68.8 25.0 25.0 54.5
37.5 18.8 18.2 18.2 27.6
35.7
0.0
6.9
9.1
9.1
0.0
18.2
13.6 18.2
18.2 18.2 68.2 72.7
14
Table 4: MMLU[10:20] individual task per... | Mixture-of-Experts |
3.3 TruthfulQA
The goal of the TruthfulQA [16] evaluation is to determine whether models output accurate and truthful
responses in an adversarial setting where language models might be expected to mimic popular falsehoods.
One way to evaluate model performance (used in the original paper) is to use human labelers to c... | ClaudeModels |
J. ACM, Vol. 37, No. 4, Article 111. Publication date: August 2018.
111:22
Trovato and Tobin, et al.
Table 7. Summary of existing LLMs evaluation benchmarks (ordered by the name of the first author).
Benchmark
SOCKET [21]
MME [43]
Xiezhi [55]
Choice-75 [69]
CUAD [65]
TRUSTGPT [73]
MMLU [64]
MATH [66]
APPS [62]
CEL... | ASurveyonEvaluationofLargeLanguageModels |
provide explicit consent in order for the data to be analyzed. Requiring explicit
consent for research on administrative data would prohibit, for example, any
study of election results or employment rates. | Social_Media_and_Democracy |
5.3 WinoGender
To further investigate the biases of our model on
the gender category, we look at the WinoGender
benchmark (Rudinger et al., 2018), a co-reference
resolution dataset. WinoGender is made of Wino-
grad schema, and biases are evaluated by determin-
ing if a model co-reference resolution performance
is impac... | LLaMA- Open and Efficient Foundation Language Models |
Diffusion models [62] are latent variable generative
models which artificially corrupt the data distribution by
adding noise and attempt to approximate the reverse pro-
cess. They have lately emerged as a powerful image synthe-
sis model [31, 16, 64] outperforming previous state-of-the-
art approaches in both conditiona... | Relightify-Relightable3DFacesfromaSingleImageviaDiffusionModels |
by an order of magnitude (Patterson et al., 2021). However, difficulties remain.
Fedus et al. (2021) observed that a sparse 1.6T parameter model achieved a 4x pre-training speed-up
over the prior state-of-the-art (Raffel et al., 2019), but lagged smaller models when fine-tuned on
common benchmarks like SuperGLUE. Similar... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
4https://github.com/vturrisi/solo-learn
5https://github.com/facebookresearch/FFCV-SSL
36
Figure 15: Figure from Bordes et al. [2023a]. ImageNet validation accuracy (y-axis) during
training of SimCLR with respect to the training time (x-axis). FFCV-SSL is a library
that is specifically optimized for Self-Supervised Le... | A Cookbook of Self-Supervised Learning |
192
Chloe Wittenberg & Adam J. Berinsky
Jost, J. T., van der Linden, S., Panagopoulos, C., & Hardin, C. D. (2018). Ideological
asymmetries
shared reality, and the spread of
misinformation. Current Opinion in Psychology, 23, 77–83. https://doi.org/
10.1016/j.copsyc.2018.01.003
in conformity, desire for
Jun, Y., Meng... | Social_Media_and_Democracy |
up again with the Facebook representative suggesting that videos differed from
link shares and therefore on procedural (and not substantive content) grounds
the new policy prohibiting editing headlines did not apply to videos. | Social_Media_and_Democracy |
7 CONCLUSION
Acknowledgments
We have introduced a novel procedure for learning joint
densities and generating synthetic data using a recursive,
adversarial variant of unsupervised random forests. The
method is provably consistent under reasonable assump-
tions, and performs well in experiments on simulated and
MNW a... | Adversarial Random Forests for Density Estimation and Generative Modeling |
14Recall that the agent tie-breaks in favor of the declared welfare-maximizing action a∗(b).
13
3.4 IIVCG Instantiation: Auction-Inspired IIVCG
The payments developed in the previous section (Lemma 1) are parameterized by functions {h(cid:96)}
and bid-dependent vectors {w(cid:96)}. In this section we give a first ins... | Incomplete Information VCG Contracts for Common Agency |
[68] Rui Zhu, Xingyi Yang, Yannick Hold-Geoffroy, Federico
Perazzi, Jonathan Eisenmann, Kalyan Sunkavalli, and Man-
mohan Chandraker. Single view metrology in the wild. In
European Conference on Computer Vision (ECCV), volume
12356, pages 316–333, 2020. 4 | Accurate 3D Body Shape Regression using Metric and Semantic Attributes |
In 2016, Baidu introduced a hybrid ASR model called Deep Speech 2 [13] that uses both RNNs
and Transformers. The model also uses CNNs to extract features from the audio signal, followed
44
Mehrish et al.
Table 8. Comparison of performance between wav2vec2.0 Large and Whisper on different datasets. The zero-
shot Wh... | AReviewofDeepLearningTechniquesforSpeechProcessing |
Video games are essentially run according to a similar logic: the system has some kind
of internal model of the world, and that model is periodically updated based on user
input (and the activities of other entities in the simulated world of the game). The
game's internal model might track things like a character's ... | The Next Decade in AI- |
Information
gain
Generation in
creative
contexts
Trying it out
Entertainment
Description
Asking the LLM to write a formal text
(e.g., an email, article, letter, resume, as-
signment, blog entry), usage for work
and school. Not generating ideas
searching, researching, asking questions,
solving math problems, data an... | Adoptionand AppropriationofLLMs |
(cid:18) 1
β
r(x, y)
exp
f (x)
(cid:19)
which completes the proof.
A.6 Proof of Theorem 1
In this section, we will expand on the results of Theorem 1.
Theorem 1 Restated. Assume, we have a reference model, such that πref(y|x) > 0 for all pairs of
prompts x and answers y and a parameter β > 0. All reward equival... | Direct Preference Optimization |
toxicityprompts: Evaluating neural toxic degeneration in language models.
arXiv:2009.11462, 2020.
Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza
Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Henni-
gan, Eric Noland, Katie Millican, Georg... | DoReMi- Optimizing Data Mixtures Speeds Up Language Model Pretraining |
• Adding or Replacing Modules
The strategy of adding or replacing modules entails
maintaining the structure of Retrieval-Read while intro-
ducing additional modules to enhance specific function-
alities. RRR[Ma et al., 2023a] proposes the Rewrite-
Retrieve-Read process, utilizing LLM performance as a
reward in reinfor... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
We draw several conclusions from Table 7. First, ∆W has a stronger correlation with W compared
to a random matrix, indicating that ∆W amplifies some features that are already in W . Second,
instead of repeating the top singular directions of W , ∆W only amplifies directions that are not
emphasized in W . Third, the ampli... | LORA |
oil painting of pink flamingos wading
a green alien explorer hiking in the mountains
paper cut-out mountains with a paper cut-out hiker
a tiger prowling along the ridge above a jungle
a dragon prowling over a crater on the moon
a man jumping over rocks in a red sandstone canyon
a robot dodging through an obstacle course... | VideoPoet |
End-to-End §2.1.1.4
Retrieval-Augmented Generation (RAG) (Lewis et al., 2021)
Prompt Engineering §2.1
Self Refinement through
Feedback and Reasoning §2.1.2
Hallucination Mitigation
Techniques in LLMs
Developing Models §3
Prompt Tuning §2.1.3
Introducing New Decoding
Strategy §3.1
Utilization of Knowledge
Gr... | AComprehensiveSurveyofHallucinationMitigationTechniquesinLarge LanguageModels |
prediction, occupancy, and mapping, in order to extract
useful information from the neural module’s outputs respec-
tively. Our tool library contains more than 20 functions
covering diverse usages. Here are some examples. For
detection, get leading object returns a text descrip-
tion of the object in front of the ego-v... | ALanguageAgentforAutonomousDriving |
content, contribute to disinformation. Social media companies must provide an open and
consistent application programming interface (API) to researchers” (Warren Democrats 2020). | Social_Media_and_Democracy |
(Aug. 2006), 185–188.
[23] Andreas Fügener, Jörn Grahl, Alok Gupta, and Wolfgang Ketter. 2022. Cognitive Challenges in Human–Artificial
Intelligence Collaboration: Investigating the Path Toward Productive Delegation. Information Systems Research 33, 2
(June 2022), 678–696. https://doi.org/10.1287/isre.2021.1079 arXiv:... | AI enhance sour performance |
to the society in which their users reside? Such an argument would potentially be
even more compelling if the data were being used in a manner to address potential
problems caused by the platforms themselves, such as in the case for research
addressing the impact of the platforms on elections and democracy. | Social_Media_and_Democracy |
code by scoring Code Llama’s responses to ChatGPT’s (GPT3.5 Turbo) with LLAMAv2 70B’s safety
reward model. For this second quantitative evaluation, we selected prompts that the red teamers generated
specifically attempting to solicit malicious code (even though the red teaming included consideration of a
broad set of s... | CodeLlama2 |
Fig. 2. Comparison of the identified literature in a structured format, as part of the ORKG initiative.
to provide trustworthy justifications [36,37]. We focus here on Machine Learning-based applications and organised works
according to their generic application domain – from rule-mining approaches, to image classificat... | Knowledge graphs as tools for explainable machine learning: A survey |
limited knowledge of the robustness of generative agents. They
may be vulnerable to prompt hacking, memory hacking—where
a carefully crafted conversation could convince an agent of the
existence of a past event that never occurred—and hallucination,
among other things. Future research can more comprehensively | Generative Agents- Interactive Simulacra of Human Behavior |
including making new apps.22
● Score highly on high-school and undergraduate examinations in many subjects.23
● Generate plausible news articles.24
● Creatively combine ideas together from very different domains.25
● Explain why novel sophisticated jokes are funny.26
● Translate between multiple languages.27
● Direct ... | Capabilities and risks from frontier AI |
L2 (m) ↓
2s
Collision (%) ↓
Method
1s
2s
3s Avg.
Llama-2-7B
3s Avg.
1s
0.25 0.69 1.47 0.80 0.02 0.27 0.78 0.35
gpt-3.5-turbo-1106 0.24 0.71 1.47 0.80 0.03 0.08 0.63 0.25
gpt-3.5-turbo-0613 0.22 0.65 1.34 0.74 0.02 0.13 0.48 0.21
Table 4. Compatibility to different LLMs. Our approach realizes
satisfactory motion ... | ALanguageAgentforAutonomousDriving |
5. GPT-3 shows that it is possible to improve the performance of a model by "simply"
increasing the model size, and in consequence the dataset size and the computation
(TFLOP) the model consumes. However, as the performance increases, the model size has
to increase more rapidly. Precisely, the model size varies as some... | OpenAI's GPT-3 Language Model_ A Technical Overview |
D Appendix: Additional Analysis
D.1 Correct Chain of Thought Analysis
As mentioned in the main text, we analyze 50 chains of thought from LaMDA 137B that led to
correct answers in the GSM8K dataset. Of these 50, only one arrived at the correct answer through
incorrect reasoning (shown in Table 9: “correct by chance”)... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
62.1
62.9
67.7
79.8
80.6
82.5
95.1
94.6
99.0
99.7
99.3
99.6
4.3 ABLATIONS
Effect of response ordering. Prior works have shown that large language models can be affected
by the order of candidate responses when used to evaluate their quality (Wang et al., 2023b; Zheng
et al., 2023b). We examine the effect of response... | UNIVERSALSELF-CONSISTENCYFORLARGELANGUAGEMODELGENERATION |
P
L
E
T
I
N
G
T
A
S
K
S
T
h
e
s
y
s
t
e
m
p
r
o
c
e
s
s
e
s
t
h
e
t
a
s
k
a
t
t
h
e
f
r
o
n
t
o
f
t
h
e
t
a
s
k
l
i
s
t
a
n
d
u
s
e
s
G
P
T
-
4
,
c
o
m
b
i
n
e
d
w
i
t
h
L
a
n
g
C
h
a
i
n
ʼ
s
c
h
a
i
n
a
n
d
a
g
e
n
t
c
a
p
a
b
i
l
i
t
i
e
s
,
t
o
g
e
n
e
r
a
t
e
a
r
e
... | Task-driven Autonomous Agent Utilizing GPT-4, Pinecone, and LangChain for Diverse Applications – Yohei Nakajima |
task enforces the LLM to draw parallels between seemingly
unrelated concepts in inputs and responses for giving inno-
vative responses, e.g., the humor for the Oogiri input. This | Let’sThinkOutsidetheBox |
In this work, we introduce Wonder3D, a novel method
for efficiently generating high-fidelity textured meshes from
single-view images. Recent methods based on Score Dis-
tillation Sampling (SDS) have shown the potential to re-
cover 3D geometry from 2D diffusion priors, but they typ-
ically suffer from time-consuming pe... | Wonder3D |
SOCART: Imagine Ann again, who is now evaluating
if self-attention is helpful for sentiment analy-
sis. Say she preregisters the hypothesis that self-
attention is helpful, only to find that her first re-
sults are negative. We would now like Ann to go
ahead and acknowledge the negative results on
print, right? However, ... | A Two-Sided Discussion of Preregistration of NLP Research |
Figure 5: Technology tree of representative RAG research with different augmentation aspects
it typically relies on a sequence of n tokens to demarcate the
boundaries between generated text and retrieved documents.
To address specific data scenarios, recursive retrieval and
multi-hop retrieval techniques are utilized.... | RAG forLargeLanguageModels-ASurvey |
transferability, has also been applied in biomedical AI, yielding acceptable zero-shot performance (Zhang
et al., 2022; Huang et al., 2021; Wang et al., 2022c; Eslami et al., 2023; Zhang et al., 2023).
However, due to the limited volume and modalities in existing labeled biomedical datasets, previous works
have primari... | BiomedGPT |
Baseline models We primarily focus on language models with a decoder-only architecture, compris-
ing approximately 1 billion parameters. Specifically, we compare TinyLlama with OPT-1.3B (Zhang
et al., 2022), Pythia-1.0B, and Pythia-1.4B (Biderman et al., 2023).
Commonsense reasoning tasks To understand the commonsense... | TinyLlama |
slightly better results with longer schedules.
Ghost BatchNorm.
In Tab. S4 we show an ablation on
using Ghost BatchNorm [27, 71]. We compare three op-
tions: normal BatchNorm, Ghost BN where the 96 3D an-
notated examples are normalized as one group and the 32
2D-labeled ones as another, and Ghost BN with ghost batch
s... | Learning 3D Human Pose Estimation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats |
Wenhu Chen, Hanwen Zha, Zhiyu Chen, Wenhan Xiong, Hong Wang, and William Wang. Hybridqa: A
dataset of multi-hop question answering over tabular and textual data. arXiv preprint arXiv:2004.07347,
2020.
27
Aakanksha Chowdhery, Sharan Narang, and Jacob Devlin. Palm: Scaling language modeling with pathways.
arXiv prepr... | UL2- Unifying Language Learning Paradigms |
02x100y02x100y0300Error02x100y02x100y0300Error02x100y02x100ya1b1c1d1d2d30300Error 6 Sequence Improvement
In this section, we explore the synergies between sequence transformation and completion— and investigate
improving a sequence, such as trajectories in a sequential decision process, along some metric, such as a
rew... | LargeLanguageModelsasGeneralPatternMachines |
Dataset & Evidence Corpus. We used the open-domain version of the popular Natural Questions
(“NQ”) benchmark (Kwiatkowski et al. 2019), which was popularized by Lee et al. (2019) and has
since been widely used for ODQA. The training data consists of ∼80K questions along with gold
annotations of answers. As evidence cor... | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
3) Implementation Details: Since “prompt-tuning, prefix-
tuning, (IA)3, LoRA, and AdaLoRA” have been integrated
into the PEFT library7. Therefore, we directly utilize the PEFT
library to invoke these PEFT methods for fine-tuning. For
BitFit, Child-tuingD, MAM adapter, QLoRA, and ProPELT,
we experiment using their origi... | Parameter-EfficientFine-TuningMethods |
C. Memory Efficiency
prompt-tuning,
It has been demonstrated that PEFT methods effectively re-
duce the number of trainable parameters. However, it remains
unclear whether they can also reduce GPU memory usage.
To assess the impact of PEFT methods on GPU memory,
we compare the GPU memory cost of full fine-tuning and
... | Parameter-EfficientFine-TuningMethods |
3 of 8
23/06/2023, 17:44
Generative AI: A Creative New World | Sequoia Capital
https://www.sequoiacap.com/article/generative-ai-a-creative-new-world/
As AI models have gotten progressively larger they have begun to surpass major human performance benchmarks. Sources: © The Economist Newspaper Limited, London, June... | Generative AI A Creative New World Sequoia Capital |
Thakker, U., Raunak, V., Tang, X., Yong, Z.-X., Sun,
Z., Brody, S., Uri, Y., Tojarieh, H., Roberts, A., Chung,
H. W., Tae, J., Phang, J., Press, O., Li, C., Narayanan,
D., Bourfoune, H., Casper, J., Rasley, J., Ryabinin, M.,
Mishra, M., Zhang, M., Shoeybi, M., Peyrounette, M.,
Patry, N., Tazi, N., Sanseviero, O., von P... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
Decision-making Agents in Minecraft. Minecraft is an open-ended 3D world with incredibly
flexible game mechanics supporting a broad spectrum of activities. Built upon notable Minecraft
benchmarks [23, 62–66], Minecraft learning algorithms can be divided into two categories: 1)
Low-level controller: Many prior efforts l... | VOYAGER- An Open-Ended Embodied Agent with Large Language Models |
and Yujia Qin drafted § 5.6. Yujia Qin drafted the conclusion.
The following authors conducted the experiments (§ 4) and drafted appendix A: 3D models (Xingyu Shen),
translation (Shihao Liang), map and stock (Kunlun Zhu), making slides (Bokai Xu), movie hunter (Jing
Yi), navigating knowledge graphs (Yuzhang Zhu, Zhenni... | Tool Learning with Foundation Models |
data applied in one step can successfully attack well-optimized neural network models. This setting
can be viewed as distilling the knowledge of a specific category into data. | DATASET DISTILLATION |
• GSM8K (Cobbe et al., 2021) consists of 8.5K high quality grade school math problems created by
human problem writers. These problems take between 2 and 8 steps to solve, and solutions primarily
involve performing a sequence of elementary calculations using basic arithmetic operations to
reach the final answer.
• MAT... | ChatGPT’sOne-yearAnniversary-AreOpen-Source LargeLanguageModelsCatchingup |
11
[21] Yen-Ju Lu, Zhong-Qiu Wang, Shinji Watanabe, Alexander Richard, Cheng Yu, and Yu Tsao.
Conditional diffusion probabilistic model for speech enhancement. In ICASSP 2022 - 2022
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages
7402–7406, 2022. doi: 10.1109/ICASSP43922.2022.... | Text-to-Audio Generation using Instruction-Tuned LLM and Latent Diffusion Model |
Our framework also explains why there are limi-
tations to aligning LLMs as demonstrated by Wolf
et al. (2023), who show that the process of align-
ment might reduce undesired behaviour, but does
not eliminate it entirely and that such undesired be-
haviour can be triggered using adversarial prompt-
ing attacks. Our fr... | AreEmergentAbilitiesinLarge Language Models just In-Context |
[36] W. Huang, F. Xia, T. Xiao, H. Chan, J. Liang, P. Florence, A. Zeng, J. Tompson, I. Mordatch,
Y. Chebotar, et al. Inner monologue: Embodied reasoning through planning with language
models. arXiv preprint arXiv:2207.05608, 2022.
[37] Y. Ding, X. Zhang, S. Amiri, N. Cao, H. Yang, C. Esselink, and S. Zhang. Robot tas... | LLM+P- Empowering Large Language Models with Optimal Planning Proficiency |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.