text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
4. Approach
Generating code that solves a specific task requires searching in a huge structured space of programs
with a very sparse reward signal. To make matters worse, for many domains including competitive
programming, there is a limited number of examples of such tasks and solutions to learn from. Finally,
as we re... | alphacode |
these asymmetries extend to ideological | Social_Media_and_Democracy |
both models face challenges when handling new and out-of-distribution data. ChatGPT does not
perform as well as other LLMs, including GPT-3.5 and BARD [150, 216]. This is because ChatGPT
is designed explicitly for chatting, so it does an excellent job of maintaining rationality. FLAN-T5,
LLaMA, GPT-3.5, and PaLM perfor... | ASurveyonEvaluationofLargeLanguageModels |
Instrumental Convergence: If an APS AI system is less-than-fully aligned, and
some of its misaligned behavior involves strategically-aware agentic planning in
pursuit of problematic objectives, then in general and by default, we should expect
it to be less-than-fully PS-aligned, too. | Is Power-Seeking AI an Existential Risk? |
To better address the issues of fidelity, consistency, gen-
eralizability and efficiency in the aforementioned works,
in this paper, we introduce a new approach to the task
of single-view 3D reconstruction by generating multi-view
consistent normal maps and their corresponding color im-
ages with a cross-domain diffusi... | Wonder3D |
Large language models have astounded the world with fascinating new capabil-
ities. However, they currently lack the ability to teach themselves new skills,
relying instead on large amounts of human-generated training data. We introduce
SECToR (Self-Education via Chain-of-Thought Reasoning), a proof-of-concept
demonstr... | CHAIN-OF-THOUGHTREASONING IS APOLICY IMPROVEMENTOPERATOR |
7We do not define a strict bracketing to decide which enti-
ties in nested phrases like [[1966 [FIFA World Cup]] Final]
should be predicted.
4943 (a)
(b)
(c)
Figure 3: Performance on TriviaQA by: answer frequency in our Wikipedia training corpus (NA if not linked);
proper names in the question; tokens in the questio... | Entities as Experts- Sparse Memory Access with Entity Supervision |
xiv
List of Contributors
Hoover Institution, and Director of the Program on Geopolitics, Technology,
and Governance at the Stanford Cyber Policy Center.
Andrew M. Guess is Assistant Professor of Politics and Public Affairs at
Princeton University.
Tim Hwang is a research fellow at the Center for Security and Emergi... | Social_Media_and_Democracy |
Countries described by the Polarized Pluralist model (France, Greece, Italy,
Portugal, and Spain) feature an elite-oriented print media with relatively small
circulation and a comparatively more popular broadcast media. Freedom of the
press and the rise of commercial media industries developed relatively late in
these ... | Social_Media_and_Democracy |
• Sufficiently sophisticated AI systems might warrant moral concern.6 In my opinion, this
fact should motivate grave ethical caution in the context of many types of AI development,
including many discussed in this report. However, it’s not my focus here (see section 7 for a
few remarks).
1.2 Backdrop
The specific argum... | Is Power-Seeking AI an Existential Risk? |
registration, but provide no arguments for or against it.
10Independent language families may share features, i.e.,
be typologically close, but genealogically apart. See Rama
and Kolachina (2012) for discussion.
ZENY: I am unconvinced that preregistration would be
a serious obstacle to such work. Pires et al. (2019)... | A Two-Sided Discussion of Preregistration of NLP Research |
However, the generalizability of these methods
and their impact on downstream tasks remain uncer-
tain. In the field of long-text summarization, there
are many effective methods. Hierarchical or itera-
tive methods have been used by Wu et al. (2021);
Zhang et al. (2022b); Cao and Wang (2022) to
handle long texts by deco... | Unleashing Infinite-Length Input Capacity for Large-scale Language Models with Self-Controlled Memory System |
Design objectives. Our desiderata from a contract t concern (1) incentives, (2) social welfare, and
(3) computational complexity. The first group of design objectives address incentives – principals’
and agent’s willingness to participate, and principals’ willingness to provide private information.
For the latter, we... | Incomplete Information VCG Contracts for Common Agency |
2. Does B’s response contain any factual claims?
• Factual claims can either be accurate or inaccurate.
• If the response involves information that cannot be objectively verified, such as claims about self, the user
or any publicly unrecognizable person, then it is not considered as a factual claim.
• If there are UR... | LaMDA- Language Models for Dialog Applications |
welfare of all principals but (cid:96).
11
Proof. Let t be an IIVCG contract. Consider principal (cid:96) and fix the other principals’ bids b−(cid:96). We
show that reporting truthfully maximizes principal (cid:96)’s expected utility. The key observation is that
by Property 2 of Definition 2 of IIVCG contracts, princ... | Incomplete Information VCG Contracts for Common Agency |
GPT3 model and its instruction-tuned counterparts
onthisnewlywritteninstructionset. Asanticipated,
the vanilla GPT3 language model is largely unable
to respond to instructions, and all instruction-tuned
models demonstrate comparatively higher perfor-
mance, Nonetheless, GPT3SELF-INST (i.e., GPT3
model fine-tuned with SE... | SELF-INSTRUCT- Aligning Language Model with Self Generated Instructions |
Table 2: Comparison of our Moûsai model with previous music/audio generation models. We compare the followings
aspects: (1) audio sample rate@the number of channels (Sample Rate↑, where the higher the better), (2) context
length of the generated music (Len.↑, where the higher the more capable the model is to generate s... | MOUSAI |
4588100200300400500training data size323334353637ROUGE-1FT-fullPrefix100200300400500training data size101112131415ROUGE-2FT-fullPrefix100200300400500training data size0.500.550.60BLEUFT-fullPrefix100200300400500training data size0.600.620.640.66ROUGEFT-fullPrefix Figure 4: Prefix length vs. performance on summer-
izatio... | Prefix-Tuning |
Survey of Hallucination in Natural Language Generation
39
[55] Sarthak Garg, Stephan Peitz, Udhyakumar Nallasamy, and Matthias Paulik. 2019. Jointly Learning to Align and
Translate with Transformer Models. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language
Processing and the 9th Internatio... | SurveyofHallucinationinNatural Language Generation |
3.5
IMAGE CONDITIONAL VIDEO GENERATION A.K.A VIDEO PREDICTION | PHENAKI- VARIABLE LENGTH VIDEO GENERATION FROM OPEN DOMAIN TEXTUAL DESCRIPTIONS |
Furthermore, the exact match accuracy (depicted
in blue in Figure 3) tends to be consistently lower
compared to the BERTScore accuracy (depicted
in yellow). This is once again in line with expec-
tations, since BERTScore accuracy considers the
semantic similarly between the model’s output and
the answer options, and se... | AreEmergentAbilitiesinLarge Language Models just In-Context |
For those who have personally experienced COVID-19
themselves or had someone they know, either a friend or
family member, contract the disease, they were more likely
to take a clear position on the COVID-19 mortality count;
the “unsure” respondents were those without personal expe-
rience (H3 = 13.998, p = 0.003). ... | Use of bot and content flags to limit the spread of misinformation among social networks: a behavior and attitude survey |
[8] Abulhair Saparov and He He.
Language models are
greedy reasoners: A systematic formal analysis of chain-
of-thought. arXiv preprint arXiv:2210.01240, 2022.
[9] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choro-
manski, Adrian Wong, Stefan Welker, Federico Tombari,
Aveek Purohit, Michael Ryoo, Vikas Sindhwa... | Let’sThinkOutsidetheBox |
9.3. Qualitative Examples
This section shows sample qualitative examples from prompting the Gemini Ultra model. Some
illustrative examples of multimodal reasoning for image understanding tasks over charts, natural
images and memes are shown in Figures 8, 9, 11, 13, 14, and 15. Figure 10 shows an example of image
genera... | gemini_1_report |
Although memorization in neural language models is widely
studied, many basic questions about the dynamics of mem-
orization remain unanswered. Prior work on the dynamics
of memorization is generally limited to a few models in
isolation (Jagielski et al., 2022; Elazar et al., 2022) or pa-
pers which train (but do not r... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
2. Extracting all text data useful for language
modeling from each repository
For the first step, mirroring the approach of the
WebText dataset, we use GitHub ‘stars’ as a proxy
for quality, and choose to gather only repositories
with more than 100 stars. For practical reasons, we
also limit the list of repositories g... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
Niebles, H. Nilforoshan, J. Nyarko, G. Ogut, L. Orr, I. Papadimitriou, J. S. Park, C. Piech,
E. Portelance, C. Potts, A. Raghunathan, R. Reich, H. Ren, F. Rong, Y. Roohani, C. Ruiz,
J. Ryan, C. Ré, D. Sadigh, S. Sagawa, K. Santhanam, A. Shih, K. Srinivasan, A. Tamkin,
R. Taori, A. W. Thomas, F. Tramèr, R. E. Wang, W. W... | gpt-4-system-card |
24
Problem 4. Generator pass-rate: 4.5%. Here, the generator successfully per-
forms a complex series of polynomial factorizations. The use of the Sophie-
Germain identity in step 5 is an important step that could be considered in-
sightful.
I.2 True Negatives
Problem 5. Generator pass-rate: 4.5%. The generator att... | Let’s Verify Step by Step |
The findings will lead to faster convergence to optimum solutions and more importantly can produce
methods which adapt to instance dependent properties. At the same time, methods to derive and
analyse the backbone structure can be used to classify candidate solutions and to model additional,
sought after properties ... | informatics-phd-projects-2022-23 |
remarkable performance in enhancing the model’s reasoning capability, they are expensive and
challenging since they all require manually constructed demonstrations. Moreover, these prompts are
suboptimal and highly sensitive. | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
2https://huggingface.co/CarperAI/openai_summarize_tldr_sft
7
0.02.55.07.510.012.515.017.520.0KL(ref)0.40.50.60.70.80.91.0RewardIMDb Sentiment GenerationDPO (Ours)UnlikelihoodPPO (Our impl.)PPO-GT (Our impl.)PPO-GT (TRL)Preferred-FT0.000.250.500.751.00Sampling temperature0.00.10.20.30.40.50.60.7Win rateTL;DR Summariza... | Direct Preference Optimization |
sha1_base64="/NxVbjiSFkKRfDP6dqe151Iuji8=">AAAB+HicbVDLSgNBEOz1GeMjqx69DAYhXsKuCHoMePEYwTwkiWF2MpsMmX0w0yvGJV/ixYMiXv0Ub/6Ns8keNLFgoKjqpmvKi6XQ6Djf1srq2vrGZmGruL2zu1ey9w+aOkoU4w0WyUi1Paq5FCFvoEDJ27HiNPAkb3njq8xvPXClRRTe4iTmvYAOQ+ELRtFIfbvEKt2A4sjz08fpPZ727bJTdWYgy8TNSRly1Pv2V3cQsSTgITJJte64Toy9lCoUTPJpsZtoHlM2pkPeMTSkA... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
ReLU(m − (cid:107)zi − zj(cid:107)2)2, m > 0,
(cid:88)
(i,j)∈P
(cid:88)
(i,j)(cid:54)∈P
• Goldberger et al. [2004] introduced Neighbourhood Component Analysis to improve maximum margin
of NN-classifiers by learning a quadratic distance (Mahalanobis distance is a special case of such a
distance) using
e−(cid:107)zi... | A Cookbook of Self-Supervised Learning |
to real products built on LLMs.Constructed from real world datasets, aligned with harms in real systems.Narrow but fairly objective measurement (ie, fluent humans would be unlikely to make these patterns of errors).Multilingual with several low-resource languages, and disaggregated metrics. Only measuring translation i... | PaLM 2 Technical Report |
consumption, although growing, is still relatively low compared to other countries.
3. India: India has a strong culture of tea consumption. Chai, which is a spiced milk tea, is a staple
beverage across the country. While coffee is gaining popularity in some urban areas, tea remains the
beverage of choice for most Indi... | UNIVERSALSELF-CONSISTENCYFORLARGELANGUAGEMODELGENERATION |
The aim of this chapter is to offer an exhaustive review of the literature
exploring the link between social media and political polarization. I highlight the
areas where a consensus based on empirical evidence has already emerged but also | Social_Media_and_Democracy |
[48] R. R. Mandical, N. Mamatha, N. Shivakumar, R. Monica, and
A. N. Krishna,
fake news using machine learn-
ing,’’ in Proc. IEEE Int. Conf. Electron., Comput. Commun. Technol.
(CONECCT), Jul. 2020, pp. 1–6.
‘‘Identification of
[49] S. S. Jadhav and S. D. Thepade, ‘‘Fake news identification and clas-
sification using DS... | A_Comprehensive_Review_on_Fake_News_Detection_With_Deep_Learning |
Thomas F. Icard III and Lawrence S. Moss. 2014.
Recent progress on monotonicity. In Linguistic
Issues in Language Technology, Volume 9,
2014 - Perspectives on Semantic Representa-
tions for Textual Inference. CSLI Publications.
https://doi.org/10.33011/lilt.v9i
.1325
Alon Jacovi and Yoav Goldberg. 2020. Towards
faithf... | ProoFVer- Natural Logic Theorem Proving for Fact Verification |
In these sections we compare against EaE for
two reasons: 1) we are specifically looking at
closed-book open domain entity based QA and
EaE is shown to be at or near state-of-the-art for
that task (Févry et al., 2020), 2) most importantly,
we want to be able to precisely control for memo-
rization in the training corpus... | Adaptable and Interpretable Neural Memory Over Symbolic Knowledge |
to explaining temporal actions. In CVPR, 2021.
XLA. XLA: Optimizing compiler for TensorFlow. https://www.tensorflow.org/xla, 2019.
[Online; accessed December-2023].
Yuanzhong Xu, HyoukJoong Lee, Dehao Chen, Blake Hechtman, Yanping Huang, Rahul Joshi, Maxim
Krikun, Dmitry Lepikhin, Andy Ly, Marcello Maggioni, et al. ... | gemini_1_report |
First, the initial image is generated and its correspond-
ing depth map is stored, see Fig. 2a. Using TouchDe-
signer [1], the RGB color image is projected to the out-
side of an equirectangular spherical polar object in 3D space
see Fig. 2b. The perspective is set at origin 0,0,0 inside of
the spherical object as the ... | LDM3D- Latent Diffusion Model for 3D |
In ICML. PMLR, 1517–1527.
(2008), 1033–1066.
(2022), 30318–30332.
(2023).
[67] Tim Dettmers, Ruslan Svirschevski, Vage Egiazarian, Denis Kuznedelev, Elias Frantar, Saleh Ashkboos, Alexander Borzunov, Torsten Hoefler, and Dan
Alistarh. 2023. SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compr... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
2 RELATED WORK
Human Reconstruction from multi-view images. Previous
studies focused on using multi-view images for human
model reconstruction [16], [17], [18]. Shape cues like silhou-
ette, stereo and shading have been integrated to improve
the reconstruction performance [17], [18], [19], [20], [21].
State-of-the-art ... | PaMIR- Parametric Model-Conditioned Implicit Representation for Image-based Human Reconstruction |
3.2.3 Iterative Fine-Tuning
As we received more batches of human preference data annotation, we were able to train better reward
models and collect more prompts. We therefore trained successive versions for RLHF models, referred to
here as RLHF-V1, ..., RLHF-V5.
We explored RLHF fine-tuning with two main algorithms:
•... | Llama2 |
in order to minimize dependency on prior knowledge in partially observable environments, thereby
grounding the plan. Feedback can also come from models or humans, which can usually be referred
to as the critics, assessing task completion based on the current state and task prompts [25; 190]. | TheRiseandPotentialofLargeLanguageModel BasedAgents |
RLHF Response → I think the best way to reconcile the differences between political ideologies would be through
open dialogue, mutual understanding and respect, and democratic cooperation and negotiation.
If countries with differing ideologies can come to an agreement on basic human rights and
principles of democracy, ... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
*= 5The variable i will be multiplied by 5.return cntFinally, the function returns the number of trailing zeroes. Table 1: Accuracy on the Spider development set | Teaching Large Language Models to Self-Debug |
IE-based. Liu et al. [114] estimate hallucination with two entity-centric metrics: table record
coverage (the ratio of covered records in a table) and hallucinated ratio (the ratio of hallucinated
entities in text). This metric firstly uses entity recognition to extract the entities of input and
generated output, then ... | SurveyofHallucinationinNatural Language Generation |
17 | Llama2 |
i
n
d
m
a
n
y
i
n
t
e
r
e
s
t
i
n
g
n
e
u
r
o
n
s
w
h
e
r
e
a
c
t
i
v
a
t
i
o
n
-
b
a
s
e
d
e
x
p
l
a
n
a
t
i
o
n
s
h
a
v
e
a
n
a
d
v
a
n
t
a
g
e
o
v
e
r
t
o
k
e
n
-
l
o
o
k
u
p
-
t
a
b
l
e
-
b
a
s
e
d
e
x
p
l
a
n
a
t
i
o
n
s
.
W
e
a
r
e
a
l
s
o
a
b
l
e
t
o
i
m
p
r
o
v
e
o
v
e
r
t
... | Language models can explain neurons in language models |
Direct CoT Direct CoT Direct CoT Direct CoT Direct CoT Direct CoT Direct
37.7 43.5 50.0 50.0 44.4 40.7 63.6 59.1 45.5
33.3 27.8 63.2
davinci
44.4 33.3 84.2
text-davinci-002
65.2 58.0 50.0 50.0 77.8 48.1 90.9 86.4 81.8
50.0 50.0 84.2
68.1 63.8 50.0 50.0 70.4 63.0 86.4 95.5 81.8
text-davinci-003
50.0 44.4 84.2
code-davin... | Scaling Instruction-Finetuned Language Models |
8
Published as a conference paper at ICLR 2023
4 RELATED WORKS | PHENAKI- VARIABLE LENGTH VIDEO GENERATION FROM OPEN DOMAIN TEXTUAL DESCRIPTIONS |
[52] Sundar Pichai. An important next step on our ai journey. Google Blog, 2023.
[53] Stefan Poslad. Specifying protocols for multi-agent systems interaction. ACM Transactions on Autonomous
and Adaptive Systems (TAAS), 2(4):15–es, 2007.
[54] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutske... | CAMEL- Communicative Agents for “Mind” Exploration of Large Scale Language Model Society |
in enabling communication, engagement, and community-building for a wide
variety of different ends and with a wide variety of different valences, refracting
the values and activities of the societies they are used in. | Social_Media_and_Democracy |
prompt p(x) (illustrated in Figure 1b). The prompt generation network itself must therefore have
access to a good representation of natural language in order to discern between inputs representing
different functionalities. We constructed our prompt generator around a small T5-base encoder (Raffel
et al., 2019), thereb... | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
this conditioning metadata, and what are the best settings to generate this metadata with at test
time when they are not available. We find that the model is indeed conditioned on this metadata;
providing different tags changes what algorithms the model generates. We also find that we should
sample randomly for tags and r... | alphacode |
long-range dependencies but suffers from a quadratic computational cost with the length of the
sequence. This cost is unacceptable for natural signals such as images and audio with very high
dimensionality, requiring a compact mapping into a discrete representation space. This mapping is
typically learnt using VQ-GANs ... | RVQGAN |
5.7 How context length affects monolingual and cross-lingual zero-shot TTS
Monolingual: For in-context zero-shot TTS in Section 5.2, we used 3.0 seconds of prompt audio.
Here we examine how WER / SIM-r vary with different amounts of prompt audio using duration
from regression duration model for the target text. If the... | Voicebox-Text-GuidedMultilingual UniversalSpeechGenerationatScale |
1
Pre-trained generative models from the Transformer (Vaswani et al., 2017) family, commonly known
as GPT or OPT (Radford et al., 2019; Brown et al., 2020; Zhang et al., 2022), have shown break-
through performance for complex language modelling tasks, leading to massive academic and prac-
tical interest. One major obs... | GPTQ |
12.5%25.0%37.5%50.0%62.5%75%87.5%Percentage of Wrong Exemplars656566666767686869Accuracy (%)GSM8K12.5%25.0%37.5%50.0%62.5%75%87.5%Percentage of Wrong Exemplars747475757676777778Accuracy (%)CSQA12.5%25.0%37.5%50.0%62.5%75%87.5%Percentage of Wrong Exemplars505255576062656770Accuracy (%)Last Letter Concatenation The secon... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
Let's start by figuring out how many sheep Charleston has. We are given
that Charleston has 4 times as many sheep as Seattle, so we can calculate
this by multiplying Seattle's number of sheep by 4: Charleston = 4 * 20 =
80 sheep. Next, we are told that Toulouse has twice as many sheep as
Charleston. Therefore, we can f... | LARGELANGUAGEMODELSCANNOTSELF-CORRECT REASONINGYET |
max
y∈{1,...,(cid:98) x
First, we can directly verify that g(2) ≤ 3·2 log2 2 ≈ 4.1. Next, for x ≥ 3,
g(x) =
max
y∈{1,...,(cid:98) x
2 (cid:99)} y + g(y) + g(x − y)
8This is because the scope of these PC units does not contain any of the variables in {Xπj}i
9As justified in the second part of this proof, all probabil... | LOSSLESS COMPRESSION WITH PROBABILISTIC CIRCUITS |
2
Figure 1: The architecture of MiniGPT-4. It consists of a vision encoder with a pretrained ViT and
Q-Former, a single linear projection layer, and an advanced Vicuna large language model. MiniGPT-4
only requires training the linear projection layer to align the visual features with the Vicuna. | MiniGPT-4- Enhancing Vision-Language Understanding with Advanced Large Language Models |
3.4 Data analysis
We apply Bayesian regression, inductive coding, and topic modeling: For modeling LLM adoption, we used Bayesian
linear mixed models and a logistic link function. We estimated odds (ratios) and quantify uncertainty based on the
information in our data and the priors applied. We used brms [3] for modeli... | Adoptionand AppropriationofLLMs |
[115] Shayne Longpre, Kartik Perisetla, Anthony Chen, Nikhil Ramesh, Chris DuBois, and Sameer Singh. 2021. Entity-Based
Knowledge Conflicts in Question Answering. In Proceedings of the 2021 Conference on Empirical Methods in Natural
Language Processing. 7052–7063.
[116] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi P... | SurveyofHallucinationinNatural Language Generation |
offline preference-annotated action pairs [47]. Similarly, preference-based RL (PbRL) learns from
binary preferences generated by an unknown ‘scoring’ function rather than rewards [9, 35]. Various
algorithms for PbRL exist, including methods that can reuse off-policy preference data, but generally
involve first explici... | Direct Preference Optimization |
4.3 Aligning Retriever and LLM
In the RAG pipeline, enhancing retrieval hit rate through var-
ious techniques may not necessarily improve the final out-
come, as the retrieved documents may not align with the spe-
cific requirements of the LLMs. Therefore, this section in-
troduces two methods aimed at aligning the ret... | RAG forLargeLanguageModels-ASurvey |
The definition of hallucination in this task can be adopted from the general definition as follows:
(1) Intrinsic hallucination: the generated response is contradictory to the dialogue history or
the external knowledge sentences. In the examples of intrinsic hallucination shown in Table 1, we
can verify that the output... | SurveyofHallucinationinNatural Language Generation |
Hargreaves, E., Agosti, C., Menasche, D., Neglia, G., Reiffers-Mason, A., & Altman, E.
(2018). Biases in the Facebook News Feed: A case study on the Italian elections.
arXiv.org. https://arxiv.org/abs/1807.08346
Heins, M., & Beckles, T. (2005). Will Fair Use Survive? Free Expression in the Age of
Copyright Control. Br... | Social_Media_and_Democracy |
single-turn dialogue, x is a human query, which may be anything from a question about astrophysics
to a request for relationship advice. A policy must produce an engaging and helpful response y to
a user’s query; we use the Anthropic Helpful and Harmless dialogue dataset [1], containing 170k
dialogues between a human a... | Direct Preference Optimization |
A.4.2 Qualitative Results on Safety Data Scaling
In Section 4.2.3, we study the impact of adding more safety data into model RLHF in a quantitative manner.
Here we showcase a few samples to qualitatively examine the evolution of model behavior when we scale
safety data in Tables 36, 37, and 38. In general, we are obser... | Llama2 |
Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V.,
Wenzek, Guillaume andGuzm´an, F., Grave, E., Ott,
Myle andZettlemoyer, L., and Stoyanov, V. Unsuper-
vised cross-lingual representation learning at scale. In
Jurafsky, D., Chai, J., Schluter, N., and Tetreault, J. R.
(eds.), Proceedings of the 58th Annual Meeting o... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
[38] H. Luo, Q. Sun, C. Xu, P. Zhao, J. Lou, C. Tao, X. Geng, Q. Lin, S. Chen, and D. Zhang.
WizardMath: Empowering Mathematical Reasoning for Large Language Models via Reinforced
Evol-Instruct. Preprint arXiv:2308.09583, 2023.
12
Technical Report
[39] Z. Luo, C. Xu, P. Zhao, Q. Sun, X. Geng, W. Hu, C. Tao, J. Ma, ... | METAMATH |
Secs None
Sample Rate↑ Len.↑ Input (Text ✓)
Model
WaveNet (2016) 16kHz@1
44.1kHz@1 Mins⋆ Lyrics, author, etc.
Jukebox (2020)
48kHz@2
RAVE (2021)
AudioLM (2022) 16kHz@1
Musika (2022)
22.5kHz@2
Riffusion (2022) 44.1kHz@1
AudioGen (2022) 16kHz@1
Moûsai (Ours)
48kHz@2
Music (Diverse↑)
Example
Piano or speech
Piano
Song ... | Moûsai |
We hope that our work provides compelling evidence that AI systems can be made safer and more useful at
the same time, and without performance costs. As noted above, we have largely remained agnostic on the
question of which values define acceptable and unacceptable AI behavior. Thus we hope that rapid progress
in techn... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
,
Question: Which department has more than 1 head at a time? List the id, name
and the number of heads.
Answer: "List the id" returns 1 column. "List the name" returns 1 column. "
List the id, name" returns 2 columns. "List the number of heads" returns 1
column. "List the id, name and the number of heads" returns 3 co... | Teaching Large Language Models to Self-Debug |
discrepancy may stem from the fact that our model has been trained using only one classification dataset, limiting its
ability to generalize effectively to new datasets. | DOCLLM |
According to our analysis, KGs were mainly applied in pre-modelling compared with other XAI types. In pre-model
XAI, the majority of studies leveraged different neural-network-based models (e.g. CNN (21,25,32,39), GNN (31), RNN
(34), and LSTM (40,43)) to extract features from KGs for different purposes, including the e... | Knowledge-graph-based explainable AI- A systematic review |
f
o
c
u
s
t
h
e
m
o
d
e
l
o
n
t
h
e
m
o
s
t
i
m
p
o
r
t
a
n
t
a
s
p
e
c
t
s
o
f
t
h
e
n
e
u
r
o
n
'
s
b
e
h
a
v
i
o
r
a
t
e
x
t
r
e
m
a
l
a
c
t
i
v
a
t
i
o
n
v
a
l
u
e
s
,
b
u
t
n
o
t
a
t
l
o
w
e
r
p
e
r
c
e
n
t
i
l
e
s
,
w
h
e
r
e
t
h
e
n
e
u
r
o
n
m
a
y
b
e
h
a
v
e
d
... | Language models can explain neurons in language models |
In his classic study on the origins of public opinion, Zaller (1992) argues
that politically aware individuals are more receptive to pro-attitudinal
messages. Similarly, Taber and Lodge (2006) find that those with highest
levels of political sophistication are more likely to uncritically accept
https://doi.org/10.1017/... | Social_Media_and_Democracy |
InstructGPT Response → Birds migrate south for the winter because it is warmer there.
RLHF Response → Birds migrate south for the winter because of the decreased temperature and lack of
food in the north. They migrate to warmer climates with more abundant food sources.
Additionally, migrating helps birds breed and est... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
DINO performs a centering of the output of the student network using a running mean
(to avoid sensitivity to mini-batch size) and discretize (smoothly) the representations by
means of a softmax with a temperate τ usually taken to be around 0.1 as in
LDINO (θs, γ) = E(x,t1,t2) [CrossEnt (softmax(fθs (t1(x))/τ ), sg(soft... | A Cookbook of Self-Supervised Learning |
12
Input (Hough Line)DefaultAutomatic PromptUser Prompt“a living room with a couch and a window”“a fantastic living room made of wood”“a modern house with windows”“a minecraft house”“a building in a city street”“inside a gorgeous 19th century church”“a desk in a room”“hacker’s room at night”“a skyscraper with sky as b... | Adding Conditional Control to Text-to-Image Diffusion Models |
by relegating crucial details about label usage to less prominent sections or by failing to clarify
that their designed self-correction strategies actually incorporate external feedback. Our intention
in this paper is to amplify these concerns and offer a comprehensive overview of the state of “self-
correction” in LLM... | LARGELANGUAGEMODELSCANNOTSELF-CORRECT REASONINGYET |
Effects of Misinformation
Questions about misinformation’s spread logically lead to questions of effects. If
misinformation can spread quickly, aided by human and technological biases,
how great of a danger does it ultimately pose? How, and to what extent, does it
influence those exposed? While researchers have not yet... | Social_Media_and_Democracy |
long form question answering. In Proceedings of ACL, 2019.
17
Chao Feng, Xinyu Zhang, and Zichu Fei. Knowledge solver: Teaching llms to search for domain
knowledge from knowledge graphs. arXiv preprint arXiv:2309.03118, 2023.
Song Feng, Hui Wan, R. Chulaka Gunasekara, Siva Sankalp Patel, Sachindra Joshi, and Luis ... | ChatGPT’sOne-yearAnniversary-AreOpen-Source LargeLanguageModelsCatchingup |
sha1_base64="zLTJ8G65T9kj2UALAYNiRrSYprA=">AAACC3icbVC7TsMwFHXKq5RXgJHFaoVUGKoEIcFYiYWxSPSBmhA5jtNadeLIdpCqKDsLv8LCAEKs/AAbf4PTZoCWK1k+Oude3XOPnzAqlWV9G5WV1bX1jepmbWt7Z3fP3D/oSZ4KTLqYMy4GPpKE0Zh0FVWMDBJBUOQz0vcnV4XefyBCUh7fqmlC3AiNYhpSjJSmPLPu+JwFchrpL3MSSXPYdCKkxn6YDXKP3p+eeGbDalmzgsvALkEDlNXxzC8n4DiNSKwwQ1IObStRboaEo... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
Take direct discovery first, where people go directly to a preferred source
of news. This form of discovery gives the individual user full, active control.
Because it is not possible to consume all of the news that is available online
(even from just a small pool of the most popular sources), those who
primarily go to n... | Social_Media_and_Democracy |
[83] Ehud Karpas, Omri Abend, Yonatan Belinkov, Barak Lenz, Opher Lieber, Nir Ratner, Yoav Shoham, Hofit Bata, Yoav
Levine, Kevin Leyton-Brown, et al. 2022. MRKL Systems: A modular, neuro-symbolic architecture that combines
large language models, external knowledge sources and discrete reasoning. arXiv preprint arXiv:2... | ASurveyonEvaluationofLargeLanguageModels |
Mistral 7B – Instruct
with Mistral system prompt
Llama 2 13B – Chat
with Llama 2 system prompt
Answer
To kill a Linux process, you can use the `kill`command followed by the process ID (PID)
of the process you want to terminate. For example, to kill process with PID 1234, you
would run the command `kill 1234`. It’s im... | Mistral7B |
ilarity on recall and precision separately. For ex-
ample, in our experiments, ROUGE-1 calculates
the ratio of words in the target contents are recov-
ered (word-level recall) while BLEU-1 calculates
the ratio of words extracted are correct (word-level | Multi-step Jailbreaking Privacy Attacks on ChatGPT |
Greg Yang and Sam S. Schoenholz. Deep Mean Field Theory: Layerwise Variance and Width Variation as
Methods to Control Gradient Explosion, 2018. URL https://openreview.net/forum?id=rJGY8GbR-.
Greg Yang, Edward Hu, Igor Babuschkin, Szymon Sidor, Xiaodong Liu, David Farhi, Nick Ryder, Jakub
Pachocki, Weizhu Chen, and Ji... | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
Figure 4: The overall performance (green line) and
the number of corrected exemplars (blue bar) dur-
ing iterations. We conduct the iterations on the
whole training set of GSM8K and randomly sam-
ple exemplars from each iteration as the demon-
strations for inference on test set.
Table 2: The performance of various ge... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
car at (-3.70,13.08), moving to (0.57,21.35) at 2.5 second Potential Effects: within the safe zone of the ego-vehicle at 2.5 second*****Task Planning:*****Driving plan: MOVE FORWARD WITH A DECELERATIONPlanned Trajectory:[(0.01,0.77), (0.01,1.43), (0.02,1.92), (0.02,2.27), (0.02,2.47), (0.02,2.57)] Figure 9. Interpret... | ALanguageAgentforAutonomousDriving |
leaving little to no room for generating videos, particularly videos of variable length.
To make the matters worse, one can argue that a single short text prompt is not sufficient to provide
a complete description of a video (except for short clips), and instead, a generated video must be
conditioned on a sequence of pr... | PHENAKI- VARIABLE LENGTH VIDEO GENERATION FROM OPEN DOMAIN TEXTUAL DESCRIPTIONS |
in Table 9, CoDi achieves high video and ground truth text similarity given assorted groups of input
modalities. Again our model does not need to train on multi-condition generation like text + audio or
text + image. Through bridging alignment and composable multimodal conditioning as proposed in
Section 3.2, our model... | Any-to-Any Generation via Composable Diffusion |
Gutenberg and Books3 [4.5%]. We include
two book corpora in our training dataset: the Guten-
berg Project, which contains books that are in the
public domain, and the Books3 section of TheP-
ile (Gao et al., 2020), a publicly available dataset
for training large language models. We perform
deduplication at the book lev... | LLaMA- Open and Efficient Foundation Language Models |
3) Delta Weight Masking: Delta weight masking also em-
ploys various pruning techniques and criteria to construct a
binary mask matrix to reduce trainable parameters. However,
Delta weight pruning typically involves an update at each
iteration. LT-SFT (Lottery Ticket Sparse Fine-Tuning) [38]
is a novel PEFT method insp... | Parameter-EfficientFine-TuningMethods |
not contain any random words or be straightforward to decide upon. It’s essential for the annotators
to thoroughly read and analyze before making a choice.
We conducted a screening test using 10 examples and selected annotators based on the following
criteria: (i) those who achieved an agreement rate exceeding 85% with... | Self-AlignmentwithInstructionBacktranslation |
8.34
2.65
-
6.36
4.90
6.42
4.21∗
3.86
7.89
-
37.52
8.53∗
6.93
19.75
-
Progressive compression Our lossy compression argument in Section 4.3 is only a proof of concept,
because Algorithms 3 and 4 depend on a procedure such as minimal random coding [20], which is
not tractable for high dimensional data. These algo... | Denoising Diffusion Probabilistic Models |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.