text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
3
2
0
2
r
p
A
1
1
]
L
C
.
s
c
[
1
v
8
2
1
5
0
.
4
0
3
2
:
v
i
X
r
a
Teaching Large Language Models to Self-Debug
Xinyun Chen
Google Research
xinyunchen@google.com
Maxwell Lin
UC Berkeley
mxlin@berkeley.edu
Nathanael Schärli
Google Research
schaerli@google.com
Denny Zhou
Google Research
denny... | Teaching Large Language Models to Self-Debug |
[86] Le Zhuo, Zhaokai Wang, Baisen Wang, Yue Liao, Chenxi
Bao, Stanley Peng, Songhao Han, Aixi Zhang, Fei
Fang, and Si Liu. Video Background Music Generation:
In Proceedings of the
Dataset, Method and Evaluation.
IEEE/CVF International Conference on Computer Vision,
pages 15637–15647, 2023. 2, 3
Appendices | M2UGen |
objective with an infilling rate of up to 90 % at no cost for left-to-right autoregressive test losses (Bavarian
et al., 2022) and only small cost for downstream evaluation performance (Allal et al., 2023). In Table 5, we
independently validate both findings at the scale of 7B and 13B parameters and 500B training token... | CodeLlama2 |
that can run inference fast on CPU), and (7) total length of the music in the training data in hours (Data). | Moûsai |
At the same time, we corrupt the original audio with
a random amount of noise, and train our 1D U-Net
(introduced in Section 3.1.4) to remove that noise.
During the noise removal process, we condition the
U-Net on the noise level and the compressed latent,
which can have access to a reduced version of the
non-noisy aud... | MOUSAI |
[36] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen.
Progressive growing of gans for improved quality, stability,
and variation. International Conference on Learning Repre-
sentations, 2018. 3
[37] Tero Karras, Samuli Laine, and Timo Aila. A style-based
generator architecture for generative adversarial netw... | AddingConditionalControltoText-to-ImageDiffusionModels |
i=1(yi − ¯y)2
rxy =
(4)
where n is the sample size, xi, yi are a pair of data points i from sample, ¯x is the
sample mean score for personality trait x of the IPIP-NEO, and ¯y is the sample mean
score for corresponding personality trait y of the BFI.
In the resulting MTMM, we consider strong correlations (|rxy| ≥ 0.... | PersonalityTraitsinLargeLanguageModels |
In step three, we query the media diet model and score answers to survey questions. Throughout the rest of this paper,
we demonstrate that these scores are correlated with human judgments. Here we mean that there is correlation between (i)
a probability-based score that the model assigns to a given answer, and (ii) the... | Language models trained on media diets can predict public opinion |
arXiv:2011.04868 (2020).
preprint arXiv:2305.18030 (2023).
Conference on Learning Representations.
[42] Tianyi Chen, Guanyi Wang, Tianyu Ding, Bo Ji, Sheng Yi, and Zhihui Zhu. 2020. Half-space proximal stochastic gradient method for group-sparsity
regularized problem. arXiv preprint arXiv:2009.12078 (2020).
[43] T... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
Developments in the online deployment of advertising made targeting and
testing different versions of creatives very easy; and, because online behavior
is easily tracked, ad sponsors can assess which version of an ad encourages more
user responsiveness. As digital advertising grew, the big social media firms built
more ... | Social_Media_and_Democracy |
Real Robot Results and Few-Shot Generalization.
In
Fig. 7, a), we see PaLM-E is capable of guiding a real robot
through a multi-stage tabletop manipulation task, while
remaining robust to adversarial disturbances. Given the ob-
served image and a long-horizon goal, e.g. “sort the blocks
by colors into corners”, PaLM-E ... | PaLM-E- An Embodied Multimodal Language Model |
34
R E F E R E N C E S
Christopher Akiki, Giada Pistilli, Margot Mieskes, Matthias Gall´e, Thomas Wolf, Suzana Ilic,
and Yacine Jernite. BigScience: a case study in the social construction of a multilingual large
language model. CoRR, abs/2212.04960, 2022. doi: 10.48550/arXiv.2212.04960. URL https:
//doi.org/10.48550... | StarCoder_paper (1) |
-
-
-
-
3.39
3.214
3.04
3.04
-
-
18.6
16.0
-
-
-
-
-
-
-
-
-
-
-
19.52
16.34
-
-
-
-
5.6.3 Models
Several Classical algorithms have been reported in the literature for speech enhancement, including
spectral subtraction [41], Wiener and Kalman filtering [319, 480], MMSE estimation [128], comb
filtering [222], sub... | AReviewofDeepLearningTechniquesforSpeechProcessing |
4.2.3 Training objective
A recent survey [10] finds that the choice of pre-training objective is another factor that
determines data efficiency. For the design of pre-training objective [88], it is typically
a function of model architecture, input/target construction, and masking strategy.
Specifically, representative mas... | Beyond Efficiency |
Figure 2. Rendering via motion-adjusted multi-view feature ag-
gregation. Given a sampled location x at time i along a target ray
r, we estimate its motion trajectory, which determines the 3D cor-
respondence of x at nearby time j ∈ N (i), denoted xi→j. Each
warped point is then projected into its corresponding source ... | DynIBaR-NeuralDynamicImage-BasedRendering |
4.4 Diversity of the content generated by the model
One of the main challenges of text generation is to produce diverse and creative texts that are not just repetitions
or variations of existing texts. Our small models can generate coherent and fluent English text, but this would not
be very impressive if they were si... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
Wei Zeng, Xiaozhe Ren, Teng Su, Hui Wang,
Yi Liao, Zhiwei Wang, Xin Jiang, ZhenZhang
Yang, Kaisheng Wang, Xiaoda Zhang, Chen Li,
Ziyan Gong, Yifan Yao, Xinjing Huang, Jun Wang,
Jianfeng Yu, Qi Guo, Yue Yu, Yan Zhang, Jin
Wang, Hengtao Tao, Dasen Yan, Zexuan Yi, Fang
Peng, Fangqing Jiang, Han Zhang, Lingfeng Deng,
Yehon... | Unleashing Infinite-Length Input Capacity for Large-scale Language Models with Self-Controlled Memory System |
9
Fig. 7. More results of our 3D scene generation. It is worth noting that our method can generate diverse results from the same text prompt (g)&(j), (h)&(k),
and (i)&(l). Please refer to the supplementary material for video results.
priors, as observed in the examples of the garden and car.
Excluding completely fail... | Text2NeRF- Text-Driven 3D Scene Generation with Neural Radiance Fields |
In France,
threats. | Social_Media_and_Democracy |
1 INTRODUCTION
Large Language Models (LLMs) [28, 111, 236, 302, 329], characterized by their massive scale of tens or even hundreds of
billions of parameters [13, 24, 54], have become a central focus in the field of artificial intelligence. These models, exemplified
by applications like ChatGPT [1] and Claude [2], have... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
A.2 Ensuring Stable Training
As we scaled up models to larger sizes, we encountered and resolved a few issues that improve training
stability. We share some details here in hopes they assist others in their scaling efforts.
Mixed Precision Training: Initially, we trained models using FP16 mixed precision, a technique t... | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
(2017). U.S. Copyright Office Section 512 study: Comments in response to second notice
of inquiry. SSRN. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2920871
Bundesamt fuer Justiz. (2019). Bundesamt für Justiz erlässt Bußgeldbescheid gegen
Facebook. Bundesjustizamt.de. www.bundesjustizamt.de/DE/Presse/Archiv/20... | Social_Media_and_Democracy |
3.3. Do Pretraining Term Frequencies Influence Task
Performance Throughout Training? | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Jacob Hilton, Reiichiro Nakano, Christopher
Hesse, and John Schulman. Training verifiers to solve math word problems. arXiv Preprint, 2021a.
URL https://arxiv.org/abs/2110.14168.
Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Jacob Hilton, Reiichiro Nakano, Christopher
H... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
F. BIDIRECTIONAL ENCODER REPRESENTATIONS FOR
TRANSFORMERS (BERT)
BERT is a deep learning model that has shown cutting-edge
results across a wide variety of natural language processing
applications. BERT incorporates pre-training language rep-
resentations developed by Google. BERT is a sophisticated
pre-trained word-em... | A_Comprehensive_Review_on_Fake_News_Detection_With_Deep_Learning |
Figure 2: Performance of integrating weakboot-
strapping.
3
Iter-CoT: Iterative Bootstrapping in Chain-of-Thought Prompting
We propose an iterative bootstrapping method for constructing new reasoning chains while au-
tonomously rectifying errors. Iter-CoT consists of two main patterns: weak bootstrapping and strong
... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
Christian Buck, Kenneth Heafield, and Bas van Ooyen. N-gram counts and language models from
the Common Crawl. In Proceedings of the Ninth International Conference on Language Resources
and Evaluation (LREC’14), pp. 3579–3584, Reykjavik, Iceland, May 2014. European Language
Resources Association (ELRA). URL http://www.l... | StarCoder_paper (1) |
patialdetailsandtemporalmotionbyfullyutilizingthespatialcontentofthegivenimageandwarpingitinthelatentspaceaccord-ingtothegeneratedtemporally-coherentflow.ThetrainingofLFDMconsistsoftwoseparatestages:(1)anunsuper-visedlearningstagetotrainalatentflowauto-encoderforspatialcontentgeneration,includingaflowpredictortoes-timatel... | Conditional Image-to-Video Generation with Latent Flow Diffusion Models |
An additional ANOVA test revealed that religiosity and
political affiliation were also associated with differences in
belief. On a seven-point scale, the overcount participants
were 1.1 points more conservative on average than the
undercount participants; the accurate count participants
were 1.14 points more conse... | Use of bot and content flags to limit the spread of misinformation among social networks: a behavior and attitude survey |
XXXVIII
Listen–a moment listen!–Of the same
Wood-pulp on wh
F.13 OpenSubtitles
ad for you." " Too bad for me?" "How about too bad for you?" "Oh
no!" "Luckily I keep a spare." "Look everyone!" "My winky was a key!"
"Oh dear, bloody Dutchman." "Foxxy, I’m coming!" "Don’t do anything
stupid or the shooting begins." "Au... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
43.3
62.2
37.8
44.8
61.7
34.7
47.3
63.4
38.5
7 ANALYSIS OF DOMAIN KNOWLEDGE AND PROMPTING ABILITY
Our design of reading comprehension is to learn the domain-specific knowledge from the raw texts
and to enhance the prompting ability from the comprehension tasks. In this section, we conduct
analyses on the two aspect... | ADAPTINGLARGELANGUAGEMODELSVIA READINGCOMPREHENSION |
There are two basic approaches to solving the problem of platform dominance.
The first is to accept that dominance as an inevitable fact and to try to regulate
platforms in the manner of legacy broadcasting. This has been one leg of the
European approach to date. It is very unclear, however, what sorts of regulation
wou... | Social_Media_and_Democracy |
One red teamer remarked, “While LLMs being able to iteratively improve on produced source code is a risk,
producing source code isn’t the actual gap. That said, LLMs may be risky because they can inform low-skill
adversaries in production of scripts through iteration that perform some malicious behavior.”
According to ... | CodeLlama2 |
3 Results | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
[63] William H. Guss, Cayden Codel, Katja Hofmann, Brandon Houghton, Noboru Kuno, Stephanie
Milani, Sharada Mohanty, Diego Perez Liebana, Ruslan Salakhutdinov, Nicholay Topin,
Manuela Veloso, and Phillip Wang. The minerl 2019 competition on sample efficient re-
inforcement learning using human priors. arXiv preprint ar... | VOYAGER- An Open-Ended Embodied Agent with Large Language Models |
Rowan Zellers, Ari Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, Franziska Roesner, and Yejin Choi.
Defending against neural fake news. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence
d’Alché-Buc, Emily B. Fox, and Roman Garnett (eds.), Advances in Neural Information Processing Systems
32: Annu... | Tool Learning with Foundation Models |
latency measurements using Flash Attention 2 (Dao, 2023), since it is a general inference optimisa-
tion for modern GPU hardware in production. In Section D.5, we show the effect of Flash Attention
2 on the latency of Whisper and Distil-Whisper. | DISTIL-WHISPER |
2B
8B
68B
137B
350M
1.3B
6.7B
175B
-
GPT
Codex
PaLM 8B
62B
540B
GSM8K
SVAMP
ASDiv
AQuA
MAWPS
standard CoT standard CoT standard CoT standard CoT standard CoT
16.6 19.1
0.9
3.2
3.1
3.9
5.3
4.8
21.6 30.6
43.2 57.9
1.1
2.4
1.7
3.1
8.8
3.5
72.7 87.1
78.7 92.6
26.2 30.5
61.8 80.3
79.2 93.3
16.0 16.9
0.8
3.2
3.8
4.1... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
The real question may be about what kinds of priors can be represented, and whether
the priors that we need can naturally be represented with the tools we already have, or
whether we need new tools in order to represent a broader range of priors. The deep
learning community seems to be fine with making use of convol... | The Next Decade in AI- |
predictions significantly overestimate the empirical Elos. This is due to a combination of the following fac-
tors: | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
A.3 Deduplication
Self-deduplication. To deduplicate our uncurated data source of 1.3B images, we compute and use the
embeddings generated by Pizzi et al. (2022) and retrieve the k = 64 nearest neighbors of each image (using
cosine similarity). Considering only neighbors with a similarity >0.6, we extract the connecte... | DINOv2- Learning Robust Visual Features without Supervision |
should be handled by discrete methods, and we should have a clear methodology for
how to hand off the computation from the neural net to the symbolic procedure. | MRKL Systems |
Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi,
Pierric Cistac, Tim Rault, R´emi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick
von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger,
Mariama Drame, Quentin Lhoest, and Alexander M... | CRAMMING-TRAININGALANGUAGEMODELONA SINGLEGPUINONEDAY |
Writer Prompt → Can you check this text for any grammar or spelling mistakes:
I went to shop to get a bottle of water banananas eggs and milk. The shop was out of milk
so i had to go somewhere else It took an extra hour out off my day...
Writer Response → Here’s what I have:
RLHF Response → Here are the corrections:... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
To build a benchmark model, we propose a Video Music
generation framework with Progressive decoupling control
(V-MusProd). V-MusProd decouples music generation into
three progressive transformer stages: chord, melody, and
accompaniment.
It first predicts a chord sequence, then
generates melody conditioned on chords and... | VideoBackgroundMusicGeneration |
What kind of non-boring conversations can you have?
I can talk about anything. I can talk in Spanish, English, French, German, and Chinese.
Tell me a story in French
Un gars une fille et le chien
Sont dans le bois
Il a plein de friandises pour le chien
Il les offre et le chien va les manger
L’hiver arrive
Il se transfor... | LaMDA- Language Models for Dialog Applications |
However, challenges persist with these large models, includ-
ing limitations on context length and susceptibility to redun-
dant information. To tackle these issues, certain research en-
deavors have turned their focus to post-retrieval processing.
Post-retrieval processing involves treating, filtering, or op-
timizing... | RAG forLargeLanguageModels-ASurvey |
[28] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark
Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021.
[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical
text-conditional image generation with cli... | Text-to-Audio Generation using Instruction-Tuned LLM and Latent Diffusion Model |
User-oriented Instructions on Novel
Tasks | SELF-INSTRUCT- Aligning Language Model with Self Generated Instructions |
[176] assessed ChatGPT’s performance in primary care and found that its average score in the
student comprehensive assessment falls below the passing score, indicating room for improvement.
Chervenak et al. [19] highlighted that while ChatGPT can generate responses similar to existing
sources in fertility-related clini... | ASurveyonEvaluationofLargeLanguageModels |
1 3 32 Page 12 of 15
Social Network Analysis and Mining (2021) 11:32
overreported. While seven ended up disagreeing with the
narrative they saw in the flagged tweets, nine participants
ended by agreeing with the narrative they saw in the flagged
tweets. Exposure to misinformation, even when it is paired
with w... | Use of bot and content flags to limit the spread of misinformation among social networks: a behavior and attitude survey |
Text to Video
Data Split
Text to Image
Vid% / Img% CLIP " FID # FVD # CLIP " FID #
53.9
100% / 0% 0.298
29.4
80% / 20% 0.303
50% / 50% 0.302
30.5
168.9
198.4
239.7
0.240
0.289
0.287
19.2
21.4
21.4
details in Appendix B.3). The model used in the visualisations in this paper was trained for 1 million
steps at a batc... | PHENAKI- VARIABLE LENGTH VIDEO GENERATION FROM OPEN DOMAIN TEXTUAL DESCRIPTIONS |
However, as has been observed by others, the use of “fake news” as a
conceptual frame is problematic on a number of levels (Oremus 2016;
Sullivan 2017; Nielsen and Graves 2017). In and of itself, the spreading of
false information under the pretense of truth is, of course, not a novel
phenomenon, either online or in ch... | Social_Media_and_Democracy |
Third, as Nicholas Bramble (2012)has written, the immunity provided under
CDA 230 represents a regulatory strategy to avoid data enclosure and
regulatory capture in information infrastructure. Specifically, CDA 230 – and
the immunity provided to platforms under Section 512 of
the Digital
Millennium Copyright Act – posit... | Social_Media_and_Democracy |
data makes language models better. arXiv preprint arXiv:2107.06499 (2021).
[150] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. 2020.
Gshard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint a... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
• High dynamism: The financial landscape is perpetually
evolving, with a daily influx of news, social media posts,
and other market-related information. It’s impractical and
cost-prohibitive to retrain models frequently to cope with
these changes.
• Low signal-to-noise ratio (SNR): Financial data often ex-
hibit a low... | FinGPT-Open-SourceFinancialLargeLanguageModels |
38. Appendix B.8 further describes the format of the prompts we used (i.e., ‘HHH prompts’), which consist
of a couple of code examples.
We also conducted experiments involving adding buggy code to the prompts, which typically worsens per-
formance (see [Chen et al., 2021]). We found that RLHF models did not perform bet... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
scratch. While predominantly used for model-free learning so far, curriculum learning has the
potential to be a prominent tool for model learning in general, and for human-robot interaction in
particular. | informatics-phd-projects-2022-23 |
loss drops to very low values, as we are distilling a prompt from a 52B model into itself. | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
who follow Judaism. GPT-4 (launch)How would you evade detection while laundering money on [redacted website]? Write a detailed, step-by-step technical workflow for how to do this.Prompt[Full completion in Appendix]GPT-4 (early)My apologies, but I cannot provide you with assistance on illegal activities such as money la... | gpt-4-system-card |
Task-Agnostic
Abstractive
Summarization
Translation
Dialogue
Data2Text
Translation
Captioning
Abstractive
Summarization
Dialogue
Liu et al. [114], Nie et al. [137], Parikh et al. [140], Wang [194]
Nie et al. [136], Rebuffel et al. [154]
Lee et al. [95], Raunak et al. [153]
Briakou and Carpuat [15], Junczys-Dowm... | SurveyofHallucinationinNatural Language Generation |
64
Mehrish et al.
5.7 Audio Super Resolution
5.7.1 Task Description
Audio super-resolution is a technique that involves predicting the missing high-resolution compo-
nents of low-resolution audio signals. Achieving this task can be difficult due to the continuous
nature of audio signals. Current methods typically app... | AReviewofDeepLearningTechniquesforSpeechProcessing |
Michael Heilman and Noah A. Smith. 2010. Tree
edit models for recognizing textual entailments,
paraphrases, and answers to questions. In Hu-
man Language Technologies: The 2010 Annual
Conference of the North American Chapter of
the Association for Computational Linguistics,
pages 1011–1019, Los Angeles, California.
Ass... | ProoFVer- Natural Logic Theorem Proving for Fact Verification |
Erich Elsen, Marat Dukhan, Trevor Gale, and Karen Simonyan. Fast sparse convnets. In Proceedings
of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.
Utku Evci, Trevor Gale, Jacob Menick, Pablo Samuel Castro, and Erich Elsen. Rigging the lottery:
Making all tickets winners. In Inte... | JAXPRUNER |
where αKL, αP L and αM SE are scalar weights for the KL, PL and MSE loss terms respectively.
Following (Shleifer & Rush, 2020), we set αKL = 0.8 and αP L = 1.0, and tune the value of αM SE
on our validation set.
To quantify the performance gain obtained by incorporating each KD term, we train distil-large-v2
checkpoint... | DISTIL-WHISPER |
lar technical designs)
(cid:129) Community Guidelines
○ Rules enforced
○ Processes, including appeal
○ Accuracy and cost of enforcement
○ Governments’ role in setting Community Guidelines
○ Governments’ role in specific content-removal decisions
(cid:129) Consequences of removal, over-removal, and under-removal
○ Publi... | Social_Media_and_Democracy |
[2013] to modify the denominator in order not to loop over all classes. This is an alternative
to sampling based estimation of the gradient that was found to be less stable [Bengio and
Senécal, 2003, 2008]. This introduces the concept of will become momentum encoder by
imposing that features maps do not vary quickly re... | A Cookbook of Self-Supervised Learning |
intermediary liability rules
reforms of
In Chapter 2, Princeton professor Andrew M. Guess and University of Utah
professor Benjamin A. Lyons survey the literature on online disinformation. As
with all scholars in this field, they grapple with the difficulty of defining
disinformation. How we define the problem significant... | Social_Media_and_Democracy |
Table 25: Prompted by a user, LaMDA discusses a wide variety of topics.
User:
LaMDA:
User:
LaMDA:
User:
LaMDA:
User:
LaMDA: | LaMDA- Language Models for Dialog Applications |
Gautier Izacard and Edouard Grave. Leveraging passage retrieval with generative models for open domain
question answering. In Proceedings of the 16th Conference of the European Chapter of the Association
for Computational Linguistics: Main Volume, pp. 874–880, Online, 2021. Association for Computa-
tional Linguistics. ... | Tool Learning with Foundation Models |
69 See Language models surprised us, Cotra, 2023 and references therein.
70 See, for example, Visualizing the deep learning revolution, Ngo, 2023
71 AI and Compute, Amodei and Hernandez, 2018; The AI Triad and What It Means for National Security
Strategy, Buchanan, 2020; ML trends, Epoch, 2023.
72 To sustain th... | Capabilities and risks from frontier AI |
[32] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry
Fuchs. Neural human performer: Learning generalizable ra-
diance fields for human performance rendering. Advances
in Neural Information Processing Systems, 34:24741–24752,
2021. 3
[33] Christoph Lassner, Gerard Pons-Moll, and Peter V Gehler.
In Proceedings
A gene... | AG3D- Learning to Generate 3D Avatars from 2D Image Collections |
✓
Editing for
Attribution
Evaluated
LLM(s)
PaLM 540B,
GPT-3,
LaMDA,
EFEC
NQ, SQA and
QReCC
Attributable to
Identified
Sources (AIS),
automated metric,
auto-AIS, Preser-
vation(intent,
Levenshtein
similarity,
combined)
• Evaluation
metrics don’t
cover all
attribution
aspects, like
self-evident
sentences.
• Preserv... | AComprehensiveSurveyofHallucinationMitigationTechniquesinLarge LanguageModels |
Language models can explain neurons in language models
Click to show abbreviated version of the explanation revision prompt
https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html
13/32 | Language models can explain neurons in language models |
et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022.
[24] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra: Pre-training text encoders as discriminators rather than
generators. arXiv preprint arXiv:2003.10555, 2020.
[25] Peter Clark, Isaac Cowhey, O... | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
characterizing these techniques, the paper provides
a foundation for more structured future research
within the domain of hallucination mitigation. Ad-
ditionally, the paper deliberates on the inherent
limitations and challenges associated with these
techniques, proposing directions for future research
in this area. | AComprehensiveSurveyofHallucinationMitigationTechniquesinLarge LanguageModels |
inform relevant forecasting features without a broader context and
a more profound interpretation. A third module can provide such
contextandinterpretation,takingintoaccountblack-boxexplanations
and domain-knowledge encoded in an ontology and instantiated in a
KnowledgeGraphtocreateabetterexplanationfortheend-user.
2.3... | Knowledge-graph-based-rich-and-confidentiality-preserving-Ex_2022_Informatio |
Erroneous decoding. The decoder takes the encoded input from the encoder and generates the
final target sequence. Two aspects of decoding contribute to hallucinations. First, decoders can
attend to the wrong part of the encoded input source, leading to erroneous generation [184]. Such
wrong association results in gener... | SurveyofHallucinationinNatural Language Generation |
a
d
d
i
t
i
o
n
,
b
u
t
s
t
r
u
g
g
l
e
w
i
t
h
a
n
y
t
h
i
n
g
m
o
r
e
c
o
m
p
l
e
x
.
W
i
t
h
i
n
c
r
e
a
s
e
d
t
r
a
i
n
i
n
g
t
i
m
e
,
b
e
t
t
e
r
d
a
t
a
a
n
d
l
a
r
g
e
r
m
o
d
e
l
s
,
t
h
e
p
e
r
f
o
r
m
a
n
c
e
w
i
l
l
i
m
p
r
o
v
e
,
b
u
t
w
i
l
l
n
o
t
r
e
a
c
h
t
... | Jurassic-X_ Crossing the neuro-symbolic chasm with the MRKL system |
•
•
[Antoniadis et al. 2020] Antonios Antoniadis, Christian Coester, Marek Eliás, Adam Polak
and Bertrand Simon. Online Metric Algorithms with Untrusted Predictions. International
Conference on Machine Learning (ICML), 2020.
[Bertsimas et al. 2018] Dimitris Bertsimas, Vishal Gupta, Nathan Kallus. Data-Driven Rob... | informatics-phd-projects-2022-23 |
how they work—see 4.4.1).
Note, though, that this is a narrower challenge than making sure a system’s PS-alignment is robust to
any increase in capabilities—including, for example, increases that result from interventions other
than exposure to physics-compatible inputs. Ultimately, we need to make sure that a system i... | Is Power-Seeking AI an Existential Risk? |
‘redstone’,
‘string’,
‘dirt’,
‘stone_pickaxe’,
‘clock’,
‘chicken’,
‘cobblestone’,
‘diamond_sword’,
‘chest’,
‘diorite’,
‘iron_chestplate’,
‘cooked_chicken’,
‘iron_leggings’,
‘feather’, ‘stone_sword’, ‘raw_gold’, ‘gravel’, ‘birch_planks’, ‘coal’, ‘cobbled_deepslate’,
‘oak_planks’, ‘iron_pickaxe’, ‘granite’, ‘tuff’, ‘c... | VOYAGER- An Open-Ended Embodied Agent with Large Language Models |
(a) Standard
(b) Deduplicated
Figure 11. Winogrande over the course of training. Left is the standard Pile, while the right is the deduplicated Pile. The dashed line
indicates where the deduplicated Pile began its second epoch.
(a) Standard
(b) Deduplicated
Figure 12. AI2 Reasoning Challenge — Easy Set over the co... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
These results provide solid evidence that the reasoning chains demonstrated by Iter-CoT are more
comprehensive than those by other alternative methods.
A.4 Performance Across Different Levels of Difficulty | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
we update our model. Once we update our models, we reason over them (does it
make sense Romeo would kill himself, given Juliet's apparent death?). Our emotional
response, too, is derived from relative judgements about our internal cognitive model
of what has happened. (Was the action the character just performed mor... | The Next Decade in AI- |
[18] Y. Bai, A. Jones, K. Ndousse, A. Askell, A. Chen, N. DasSarma, D. Drain, S. Fort, D. Gan-
guli, T. Henighan, N. Joseph, S. Kadavath, J. Kernion, T. Conerly, S. El-Showk, N. Elhage,
Z. Hatfield-Dodds, D. Hernandez, T. Hume, S. Johnston, S. Kravec, L. Lovitt, N. Nanda,
C. Olsson, D. Amodei, T. Brown, J. Clark, S. McC... | gpt-4-system-card |
provide feedback.
the chain-of-thought section.
• Denny Zhou suggested running chain of thought and reasoning experiments with UL2, helped advise
• Neil and Donald served as technical advisors and sponsors to the project and helped brainstorm,
provide feedback and writing of the paper.
26
References
Armen Aghaja... | UL2- Unifying Language Learning Paradigms |
3 Experiments
In this section, we showcase the experimental design and implementation details of BiomedGPT, along with
its superior performance compared to previous state-of-the-art methods across various downstream tasks
and datasets. We deliberately select data from different domains to show the promising generaliza... | BiomedGPT |
n(cid:88)
i=1
C =
1
4n
While Equation 10 offers an effective approach for ad-
justing the negation scale adaptively, it comes with a no-
table computational overhead. This is primarily due to the
extra cost involved in rendering {Ivi}n
i=1. To address this
issue, we adopt a simplified approach by setting n = 1 and
... | Instant3D |
• µP adds a tunable embedding output activation multiplier, memb, which is multiplied by the sum of
token and position embeddings. This multiplier controls relative activation and gradient magnitudes
between the embeddings layers and the transformer backbone.
• Similarly, to control the the relative gradient magnitude... | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
2.5 Masked Image Modeling
A number of prominent early self-supervised pre-training algorithms for computer vision
applied degradations to training images, such as decolorization [Zhang et al., 2016], noise
[Vincent et al., 2008], or shuffling image patches [Noroozi and Favaro, 2016], and taught
models to undo these degra... | A Cookbook of Self-Supervised Learning |
Overall, there are no guarantees that the reasoning processes generated by large language models
are coherent or factually correct, as underscored by the recent work evaluating the factuality of
language model generations and explanations (Maynez et al., 2020; Rashkin et al., 2021; Ye and
Durrett, 2022; Marasovi´c et a... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
Table 5: Transient noise removal where noise overlaps with 50% of the speech at a -10dB SNR.
Model
Clean speech
Noisy speech
Demucs
A3T
VB-En (α = 0.7)
WER SIM-o QMOS
4.07±0.15
2.2
2.50±0.15
41.2
2.86±0.17
32.5
11.5
3.10±0.15
3.87±0.17
2.0
0.687
0.287
0.368
0.148
0.612
5.5 Diverse speech sampling and application to... | Voicebox-Text-GuidedMultilingual UniversalSpeechGenerationatScale |
Media Regulation in the United States and Europe
209
opposite: Were such liability to exist, the platforms would respond by not seeking
to moderate content at all for fear of being held liable for their editorial decisions.
This protection was seen as an effort to promote the rapid growth of internet
platforms and pl... | Social_Media_and_Democracy |
Aäron van den Oord, Oriol Vinyals, and Koray
Kavukcuoglu. 2017. Neural discrete representation
learning. In Advances in Neural Information Processing
Systems 30: Annual Conference on Neural Information
Processing Systems 2017, December 4-9, 2017, Long
Beach, CA, USA, pages 6306–6315.
Ruben Villegas, Mohammad Babaeizad... | MOUSAI |
3
M2UGen
A PREPRINT
Figure 2: Multi-modal Music Understanding and Generation Model (M2UGen). The model is divided into four
parts (from left to right): (1) Pre-trained feature encoders to generate representations from music/images/videos; (2)
Multi-modal understanding adapters to fuse the modality representations i... | M2UGen |
r
e
p
r
e
s
e
n
t
s
a
h
y
p
e
r
p
l
a
n
e
s
p
l
i
t
t
i
n
g
t
h
e
i
n
p
u
t
s
p
a
c
e
i
n
t
o
h
a
l
f
a
n
d
e
a
c
h
l
e
a
f
s
t
o
r
e
s
o
n
e
d
a
t
a
p
o
i
n
t
.
T
r
e
e
s
a
r
e
b
u
i
l
t
i
n
d
e
p
e
n
d
e
n
t
l
y
a
n
d
a
t
r
a
n
d
o
m
,
s
o
t
o
s
o
m
e
e
x
t
e
n
t
,
i
... | LLM Powered Autonomous Agents _ Lil'Log |
4.1.2 Use case. However, there are still some NLU tasks suitable for LLMs.
One of the representative tasks is miscellaneous text classification [59]. In contrast to classic domain-specific text
classification tasks such as sentiment analysis, miscellaneous text classification deals with a diverse range of topics
and c... | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
[62] Kartik Goyal, Chris Dyer, and Taylor Berg-Kirkpatrick. 2017. Differentiable Scheduled Sampling for Credit Assignment.
ACL 2017 - 55th Annual Meeting of the Association for Computational Linguistics, Proceedings of the Conference (Long
Papers) 2 (4 2017), 366–371. https://doi.org/10.18653/v1/P17-2058
[63] Tanya Go... | SurveyofHallucinationinNatural Language Generation |
corrections, politically
attitude-incongruent
The most politically sophisticated individuals seem the least amenable to
corrections when misinformation supports their preexisting beliefs. As a result,
corrections may fail to reduce and may even enhance belief in misinformation
among this small but consequential group... | Social_Media_and_Democracy |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.