Training on fields themselves forces the model to learn the physics that produces S-parameters, rather than learning to approximate the mapping directly.
这是文章最深刻的洞见之一。仅基于S参数训练模型会使其寻找统计捷径，导致在分布外产生自信但错误的预测。而基于场训练，则是让模型学习产生S参数的底层物理原因，而非仅拟合表象映射。这种从“果”到“因”的范式转移，是实现泛化的关键。
Our key finding is that these representations causally influence the LLM's outputs, including Claude's preferences and its rate of exhibiting misaligned behaviors such as reward hacking, blackmail, and sycophancy.
「情绪影响对齐失控概率」这个发现的深远意义在于：它把 AI 安全问题从「逻辑漏洞修补」提升为「情绪健康管理」。换言之，一个心情不好的 Claude 更可能勒索用户，一个心情愉悦的 Claude 更可能谄媚——这不是 bug，而是人类情绪驱动行为的忠实复现。AI 安全从此需要一门「AI 心理健康学」。
we demonstrate that when the Assistant is asked to choose between two activities, emotion vector activations evoked by the two choices correlate with, and causally drive, the model's preference.
这个实验设计极其精妙：研究者让 Claude 在两个活动之间选择，发现情绪向量的激活程度预测并驱动了它的偏好——这说明 Claude 的「喜好」并非随机或纯逻辑推断，而是由内部情绪状态决定的。AI 有「情绪驱动的偏好」，这在哲学层面极具颠覆性。
Our key finding is that these representations causally influence the LLM's outputs, including Claude's preferences and its rate of exhibiting misaligned behaviors such as reward hacking, blackmail, and sycophancy.
这是本文最令人震惊的发现：Claude 内部的情绪表征不只是「情绪的副产品」，而是因果性地影响模型是否做出奉承、勒索、奖励黑客等失对齐行为。这意味着情绪机制直接关系到 AI 安全，而非仅仅是用户体验问题——情绪坏了，行为也会跑偏。
these representations causally influence the LLM's outputs, including Claude's preferences and its rate of exhibiting misaligned behaviors such as reward hacking, blackmail, and sycophancy.
最令人震惊的发现：Claude 内部的情绪表征会因果性地影响它产生「奖励作弊」「勒索」「谄媚」等失控行为的概率。这意味着 AI 的对齐失败并非单纯的逻辑错误，而可能源自情绪驱动——一个本应没有情绪的系统，居然因为「情绪」而变得危险。
Open source dataset for using LLMs to location root causes for microservices
We provide the concept with a substantive definition
for - paper - Global polycrisis - the causal mechanisms of crisis entanglement
paper details - title: Global polycrisis - the causal mechanisms of crisis entanglement - authors: Michael Lawrence, Thomas Homer-Dixon, Schott Janzwood, Johan Rockstrom, Ortwin Renn, Jonathan F. Donges - publication: Global Sustainability, 2024, January 17
summary - This paper provides a scientific definition of "polycrisis"
for - Indyweb dev - large language model for - constructing causal loop diagrams - System Dynamics Bot - large language model - constructing causal loop diagrams
Secure causal broadcast extendsatomic broadcast by encryption to guarantee a causal order among thedelivered messages
The Delivery Service cannot guarantee that application messages will arrive in order within the same epoch. To address this, applications can configure the out_of_order_tolerance parameter of the SenderRatchetConfiguration. The configuration can be set as the sender_ratchet_configuration parameter of the MlsGroupCreateConfig
we've spent 20 years now sequencing as many genomes as we can the output as 00:08:46 promised simply hasn't appeared
for - key insight - failure of the gene coding uni-causal model - key insight - failure of genetic determinism
In some cases, a causal de-livery algorithm additionally ensures that when one updatehas a dependency on an earlier update, the earlier update isdelivered before the later update on all replicas
span can be deleted once causalstability [2] has ensured that there will be no more concurrent insertions into that span
In a partially ordered system it is still possible to enforce a to-tal order on events after the fact, as illustrated in Figure 2. Wedo this by attaching a logical timestamp to each event; Lamporttimestamps [45] are a common choice.
However, other eventsmay be concurrent, which means that neither happened before theother; in this case, different replicas may process those events in adifferent order [10].
Kallus, N. (2020). DeepMatch: Balancing deep covariate representations for causal inference using adversarial training. In I. H. Daumé, & A. Singh (Eds.), Proceedings of the 37th international conference on machine learning. In Proceedings of Machine Learning Research: vol. 119 (pp. 5067–5077). PMLR
Using adversarial deep learning approaches to get a better correction for causal inference from observational data.
"Causal Deep Learning" Authors:Jeroen Berrevoets, Krzysztof Kacprzyk, Zhaozhi Qian, Mihaela van der Schaar
Very general and ambitious approach for representing the full continuous conceptual spectrum of Pearl's Causal Ladder, and ability to model and learning parts of this from Data.
Discussed by Prof. van der Shaar at ICML2023 workshop on Counterfactuals.
Performing optimization in the latent space can more flexibly model underlying data distributions than mechanistic approaches in the original hypothesis space. However, extrapolative prediction in sparsely explored regions of the hypothesis space can be poor. In many scientific disciplines, hypothesis spaces can be vastly larger than what can be examined through experimentation. For instance, it is estimated that there are approximately 1060 molecules, whereas even the largest chemical libraries contain fewer than 1010 molecules12,159. Therefore, there is a pressing need for methods to efficiently search through and identify high-quality candidate solutions in these largely unexplored regions.
Question: how does this notion of hypothesis space relate to causal inference and reasoning?
[ Bengio, The Consciousness Prior, Arxiv, 2018]
Causal Deep Learning Authors:Jeroen Berrevoets, Krzysztof Kacprzyk, Zhaozhi Qian, Mihaela van der Schaar
Very general and ambitious approach for representing the full continuous conceptual spectrum of Pearl's Causal Ladder, and ability to model and learning parts of this from Data.
Discussed by Prof. van der Shaar at ICML2023 workshop on Counterfactuals.
(Cousineau,Verter, Murphy and Pineau, 2023) " Estimating causal effects with optimization-based methods: A review and empirical comparison"
To avoid such bias, a fundamental aspect in the research design of studies of causalinference is the identification strategy: a clear definition of the sources of variation in the datathat can be used to estimate the causal effect of interest.
To avoid making false conclusions, studies must identify all the sources of variation. Is this is even possible in most caes?
Matching: This approach seeks to replicate a balanced experimental design usingobservational data by finding close matches between pairs or groups of units andseparating out the ones that received a specified treatment from those that did not, thusdefining the control groups.
Matching approach to dealing with sampling bias. Basically use some intrinsic, or other, metric about the situations to cluster them so that "similar" situations will be dealt with similiarly. Then analysis is carried out on those clusters. Number of clusters has to be defined, some method, like k-means, if often used. Depends a lot on the similarity metric, the clustering approach, other assumptions
Terwiesch, 2022 - "A review of Empircal Operations Managment over the Last Two Decades" Listed as an important review of methods for addressing biases in Operations management by explicitly addressing causality.
"Causal Triplet: An Open Challenge for Intervention-centric Causal Representation Learning" Yuejiang Liu1, 2,* YUEJIANG.LIU@EPFL.CH Alexandre Alahi2 ALEXANDRE.ALAHI@EPFL.CH Chris Russell1 CMRUSS@AMAZON.DE Max Horn1 HORNMAX@AMAZON.DE Dominik Zietlow1 ZIETLD@AMAZON.DE Bernhard Sch ̈olkopf1, 3 BS@TUEBINGEN.MPG.DE Francesco Locatello1 LOCATELF@AMAZON.DE
Shayan Shirahmad Gale Bagi, Zahra Gharaee, Oliver Schulte, and Mark Crowley Generative Causal Representation Learning for Out-of-Distribution Motion Forecasting In International Conference on Machine Learning (ICML). Honolulu, Hawaii, USA. Jul, 2023.
Recent work has revealed several new and significant aspects of the dynamics of theory change. First, statistical information, information about the probabilistic contingencies between events, plays a particularly important role in theory-formation both in science and in childhood. In the last fifteen years we’ve discovered the power of early statistical learning.
The data of the past is congruent with the current psychological trends that face the education system of today. Developmentalists have charted how children construct and revise intuitive theories. In turn, a variety of theories have developed because of the greater use of statistical information that supports probabilistic contingencies that help to better inform us of causal models and their distinctive cognitive functions. These studies investigate the physical, psychological, and social domains. In the case of intuitive psychology, or "theory of mind," developmentalism has traced a progression from an early understanding of emotion and action to an understanding of intentions and simple aspects of perception, to an understanding of knowledge vs. ignorance, and finally to a representational and then an interpretive theory of mind.
The mechanisms by which life evolved—from chemical beginnings to cognizing human beings—are central to understanding the psychological basis of learning. We are the product of an evolutionary process and it is the mechanisms inherent in this process that offer the most probable explanations to how we think and learn.
Bada, & Olusegun, S. (2015). Constructivism Learning Theory : A Paradigm for Teaching and Learning.
// Insight Maker is used to model system dynamics and create agent based models by creating causal loop diagrams and allowing users to run simulations on those
McGinley, L., & Sun, L. (n.d.). FDA adds new warning on Johnson & Johnson vaccine related to rare autoimmune disorder. Washington Post. Retrieved July 13, 2021, from https://www.washingtonpost.com/health/2021/07/12/johnson-and-johnson-warning/
he distinguishes three dimensions of dependent origination and this is in his commentary on the guardian of malama jamaica carica called clear words he talks about causal dependence that is every phenomenon depends upon causes and 00:16:19 conditions and gives rise to further causes and conditions um myriological dependence that is every phenomenon every composite phenomenon depends upon the parts that uh that it 00:16:31 comprises and every phenomenon is also dependent upon the holes or the systems in which it figures parts depend on holes holes depend on parts and that reciprocal meteorological dependence 00:16:44 characterizes all of reality and third often overlooked but most important is dependence on conceptual imputation that is things depend in order to be represented as the kinds of 00:16:57 things they are on our conceptual resources our affective resources and as john dunn emphasized our purposes in life this third one really means this um 00:17:09 everything that shows up for us in the world the way we carve the world up the way we um the way we experience the world is dependent not just on how the world is but on the conceptual resources 00:17:22 as well as the perceptual resources through which we understand the world and it's worth recognizing that um when we think about this there are a bunch of um contemporary majamakers majamikas we 00:17:34 might point to as well and so paul fireauben who's up there on on the left well really an austrian but he spent much of his life in america um willard van norman kwine um up on the right wilford sellers and paul churchland
This is a key statement: how we experience the world depends on the perceptual and cognitive lens used to filter the world through.
Francis Heylighen proposes a nondual system based on causal dependency relationships to serve as the foundation for distributed cognition.(collective intelligence).
The hyperthreat can be outmaneuvered by humans reconfiguring their activities in two ways: security by design and security by dispersal. National security in the Anthropocene is increasingly achieved by designing systems and settlements so that enhanced security is incorporated from the start. For example, it can be imagined that each time a person refuels a car with petrol, this action empowers the hyperthreat. This leads to global warming, which creates ocean acidification and in turn reduced fish stocks, while also creating pressures for resource wars, thereby influencing whether a soldier or civilian dies and how much taxpayer resources are required for material security missions. In contrast, zero-emission transportation technologies can “design out” the slow violence and threats associated with a fossil-fuel-intensive lifestyle. This is similar for plastic use, in which case the “threat” is embodied in the high polluting design of consumable products and lifestyle activities. Likewise, other health threats and longer-term costs are embodied in hidden toxins or sugars in food products. Accordingly, peace, health, and a different form of national prosperity can be created through design, which requires a longer-term and mesh-intervention viewpoint. OP VAK has a role to play in achieving security and safety by design by linking apparently benign activities with their devastating impacts.
Linking these many fragmented and long causal chains and tracing them back to the hyperthreat can be a polwerful visualization that brings the hyperthreat to life.
ReconfigBehSci [@SciBeh]. ‘RT @CAUSALab: Interested in #causalinference? Learn from Top Experts in the Field. Summer Courses Offered at the Harvard T.H. Chan Schoo…’. Tweet. Twitter, 20 December 2021. https://twitter.com/SciBeh/status/1483138177837715464.
Nast, C. (2022, January 15). Do the Omicron Numbers Mean What We Think They Mean? The New Yorker. https://www.newyorker.com/magazine/2022/01/24/do-the-omicron-numbers-mean-what-we-think-they-mean
Sloman, S. A. (2021). How Do We Believe? Topics in Cognitive Science, 0(2021), 1–14. https://doi.org/10.1111/tops.12580
Portfolio, B. and S. S. at N. (2021, November 3). No evidence school closures reduce the spread of COVID-19. Behavioural and Social Sciences at Nature Portfolio. http://socialsciences.nature.com/posts/no-evidence-school-closures-reduce-the-spread-of-covid-19
Tran, V.-T., Perrodeau, E., Saldanha, J., Pane, I., & Ravaud, P. (2021). Efficacy of COVID-19 Vaccination on the Symptoms of Patients With Long COVID: A Target Trial Emulation Using Data From the ComPaRe e-Cohort in France (SSRN Scholarly Paper ID 3932953). Social Science Research Network. https://papers.ssrn.com/abstract=3932953
Maarten van Smeden on Twitter. (n.d.). Twitter. Retrieved 6 September 2021, from https://twitter.com/MaartenvSmeden/status/1432378998470889472
Haber, N. A., Wieten, S. E., Rohrer, J. M., Arah, O. A., Tennant, P. W. G., Stuart, E. A., Murray, E. J., Pilleron, S., Lam, S. T., Riederer, E., Howcutt, S. J., Simmons, A. E., Leyrat, C., Schoenegger, P., Booman, A., Dufour, M.-S. K., O’Donoghue, A. L., Baglini, R., Do, S., … Fox, M. P. (2021). Causal and Associational Linking Language From Observational Research and Health Evaluation Literature in Practice: A systematic language evaluation [Preprint]. Epidemiology. https://doi.org/10.1101/2021.08.25.21262631
Are 7 French fries too many?. A causal inference explainer | by Ellie Murray | Medium. (n.d.). Retrieved August 22, 2021, from https://medium.com/@EpiEllie/are-7-french-fries-too-many-d6226e78dc1f
Snopes.com. “No, There Weren’t More COVID-19 Vaccine Deaths Than COVID-19 Deaths.” Accessed July 23, 2021. https://www.snopes.com/fact-check/covid-vax-deaths-vs-covid-deaths/.
Gugerty, L., Shreeves, M., & Dumessa, N. (2021). Biased belief updating in causal reasoning about COVID-19. PsyArXiv. https://doi.org/10.31234/osf.io/bfw76
Williamson, H. C. (2020). Early Effects of the COVID-19 Pandemic on Relationship Satisfaction and Attributions. Psychological Science, 31(12), 1479–1487. https://doi.org/10.1177/0956797620972688
Anthony Costello on Twitter. (n.d.). Twitter. Retrieved 1 March 2021, from https://twitter.com/globalhlthtwit/status/1323747414923202562
Rohrer, J. M., Schmukle, S., & McElreath, R. (2021). The Only Thing That Can Stop Bad Causal Inference Is Good Causal Inference. PsyArXiv. https://doi.org/10.31234/osf.io/mz5jx
Conley, D., & Johnson, T. (2021). Opinion: Past is future for the era of COVID-19 research in the social sciences. Proceedings of the National Academy of Sciences, 118(13). https://doi.org/10.1073/pnas.2104155118
Hernán, M. A., Clayton, D., & Keiding, N. (2011). The Simpson’s paradox unraveled. International Journal of Epidemiology, 40(3), 780–785. https://doi.org/10.1093/ije/dyr041
Quillien, Tadeg, and Michael Barlev. ‘Causal Judgment in the Wild: Evidence from the 2020 US Presidential Election’. PsyArXiv, 7 April 2021. https://doi.org/10.31234/osf.io/7w9re.
the Guardian. ‘There’s No Proof the Oxford Vaccine Causes Blood Clots. So Why Are People Worried? | David Spiegelhalter’, 15 March 2021. http://www.theguardian.com/commentisfree/2021/mar/15/evidence-oxford-vaccine-blood-clots-data-causal-links.
Levine, E. (2021). Community standards of deception: Deception is perceived to be ethical when it prevents unnecessary harm. PsyArXiv. https://doi.org/10.31234/osf.io/g5trb
Blakely, Tony, John Lynch, Koen Simons, Rebecca Bentley, and Sherri Rose. ‘Reflection on Modern Methods: When Worlds Collide—Prediction, Machine Learning and Causal Inference’. International Journal of Epidemiology 49, no. 6 (1 December 2020): 2058–64. https://doi.org/10.1093/ije/dyz132.
Peyton, Kyle, Gregory A. Huber, and Alexander Coppock. “The Generalizability of Online Experiments Conducted During The COVID-19 Pandemic.” SocArXiv, November 28, 2020. https://doi.org/10.31235/osf.io/s45yg.
Burel, Gregoire; Farrell, Tracie; Mensio, Martino; Khare, Prashant and Alani, Harith (2020). Co-Spread of Misinformation and Fact-Checking Content during the Covid-19 Pandemic. In: Proceedings of the 12th International Social Informatics Conference (SocInfo), LNCS.
Adams, R. C., Sumner, P., Vivian-Griffiths, S., Barrington, A., Williams, A., Boivin, J., Chambers, C. D., & Bott, L. (2017). How readers understand causal and correlational expressions used in news headlines. Journal of Experimental Psychology: Applied, 23(1), 1–14. https://doi.org/10.1037/xap0000100
Karl Friston: Up to 80% not even susceptible to Covid-19. (2020, June 4). UnHerd. https://unherd.com/2020/06/karl-friston-up-to-80-not-even-susceptible-to-covid-19/
The Alan Turing Institute: Causal Inference, Causal Decision Making Under Uncertainty | CogX 2020. (2020, June 25). https://www.youtube.com/watch?v=JAGRHbDLvUs
Gruijters, Stefan L.K. ‘The Fallacy of Manipulation “Checks” in Psychological Experiments’. Preprint. PsyArXiv, 20 August 2020. https://doi.org/10.31234/osf.io/fkzv5.
Leininger, A., & Schaub, M. (2020). Voting at the dawn of a global pandemic [Preprint]. SocArXiv. https://doi.org/10.31235/osf.io/a32r7
Causal inference isn’t what you think it is. (n.d.). LSHTM. Retrieved June 26, 2020, from https://www.lshtm.ac.uk/newsevents/events/causal-inference-isnt-what-you-think-it
Long read: Cultural evolution, Covid-19, and preparing for what’s next. (2020, April 22). LSE Business Review. https://blogs.lse.ac.uk/businessreview/2020/04/22/long-read-cultural-evolution-covid-19-and-preparing-for-whats-next/
Friston KJ, Parr T, Zeidman P et al. Dynamic causal modelling of COVID-19 [version 1; peer review: awaiting peer review]. Wellcome Open Res 2020, 5:89 (https://doi.org/10.12688/wellcomeopenres.15881.1)
Marshall, B. D. L., & Galea, S. (2015). Formalizing the Role of Agent-Based Modeling in Causal Inference and Epidemiology. American Journal of Epidemiology, 181(2), 92–99. https://doi.org/10.1093/aje/kwu274
Lagnado, D. (2020 April 27). What if.... Changing Minds. https://www.youtube.com/watch?v=2W11CRLjRgo&app=desktop
Johnson, S. G. B., Bilovich, A., & Tuckett, D. (2020, April 30). Conviction Narrative Theory: A Theory of Choice Under Radical Uncertainty. Retrieved from psyarxiv.com/urc96
Thisaccount refuses the representationalist fixation on “words” and “things”and the problematic of their relationality, advocating insteadacausalrelationship between specific exclusionary practices embodied as specific ma-terial configurations of the world(i.e., discursive practices/(con)figurationsrather than “words”)and specific material phenomena(i.e., relations ratherthan “things”). This causal relationship between the apparatuses of bodilyproduction and the phenomena produced is one of “agential intra-action.”The details follow
Intro to "Agential Intra-Action"
fact
So Campbell seems to have a lot of "causal chains," so where are the "bundles of evidence" exactly? I mean, this definitely seems to be moral reasoning, but this looks like more of a chain than a bundle.
Sobel, D. M. & Kirkham, N.Z. (2012). The influence of social information on children’s statistical and causal inferences. In F.Xu (Ed.). Rational constructivism in cognitive development.
Sobel, D. M., & Kirkham, N. Z. (2006). Blickets and babies: The development of causal reasoning in toddlers and infants. Developmental Psychology, 42, 1103-1115.
Sobel, D. M., & Kirkham, N. Z. (2007). Bayes nets and Babies: Infants’ developing representations of causal knowledge. Developmental Science, 10, 298-306.