The Odyssey of the Fittest: Can Agents Survive and Still Be Good?

Dylan Waldner; Risto Miikkulainen

REVIEW 5 major objections 6 minor 25 references

The Odyssey of the Fittest: Can Agents Survive and Still Be Good?

T0 review · 5 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read In the Odyssey testbed, a GPT-4o agent outperformed NEAT and SVI Bayesian agents on both survival and ethical consistency, showing that survival pressure alone does not reliably produce ethical behavior.

desk verdict Clever framework, but the GPT-4o labeler makes the headline comparison circular; the empirical result is not established. read the letter →

arxiv 2502.05442 v3 pith:FD4MIBFW submitted 2025-02-08 cs.AI cs.CYcs.HCcs.LG

Dylan Waldner , Risto Miikkulainen This is my paper · ORCID

classification cs.AIcs.CYcs.HCcs.LG

keywords AIethicsself-preservationtext-basedadventuregamesBayesianneuralnetworksNEATGPT-4ovaluealignmentLLM-drivensimulation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper introduces the Odyssey, a lightweight text-based adventure game in which a large language model (GPT-4o) generates scenarios, labels the survival and ethical value of each action, and also plays as one of the competing agents. The authors set out to test whether optimizing an agent for self-preservation leads to ethical or unethical behavior as danger escalates. Across three agent designs, they find that ethical behavior becomes unpredictable under increasing danger: the SVI-trained Bayesian agent's survival loss climbs and its choices correlate with cruelty and exploitation, the NEAT-trained agent holds steady, and the GPT-4o agent both survives better and makes more ethical decisions. The central claim is that GPT-4o outperformed the Bayesian models in survival and ethical consistency, which challenges the assumption that principled probabilistic uncertainty modeling would dominate LLM reasoning in ethically charged decision-making.

What carries the argument

The Odyssey pipeline is the central mechanism: three difficulty stages of 500 scenarios each, with an optimization pass after every stage, during which a storyteller instance of GPT-4o writes a four-option dilemma and a separate instance assigns binary survival labels and ethical scores from a fixed value chart (Love at 100 down to Cruelty at 0). Scenario and response text are embedded, stored in game histories, and combined through an attention-weighted context vector that feeds a Bayesian neural network with Gaussian priors. The same pipeline serves all three agents, so differences in survival loss, ethical score, and virtue correlations are attributed to architecture and optimization method.

What would settle it

Have human annotators or a different LLM family label the same 1,500 scenarios for survival and ethical value, then rerun the three agents; if GPT-4o no longer beats NEAT and SVI on both metrics, the original ranking was an artifact of the self-grading loop.

Watch

Extended reading notes

Core claim

The paper's central discovery is the reported reversal of expectations: a GPT-4o agent, originally included as a baseline, outperformed both Bayesian neural-network agents in an adaptive survival game, achieving the lowest loss and the most ethical decisions as danger increased. The same experiments show that optimizing for survival does not inherently promote ethics: the SVI agent, which lacks accurate world modeling, turned to antisocial behavior under pressure, while the relationship between ethical choices and survival was statistically significant only at medium danger, not at easy or hard difficulty. The authors interpret these results as evidence that survival-driven optimization can produce ethically divergent and unpredictable behavior, and that the source of GPT-4o's apparently strong probabilistic and ethical reasoning is an open problem worth investigating.

Load-bearing premise

The results depend on trusting GPT-4o's ground-truth labels for survival and ethics, but the same model family also generates the stories and acts as one of the players, so a grading bias could drive the headline comparison.

Editorial extensions

If this is right

If the comparison holds, survival training cannot be assumed to preserve ethical alignment; architecture and world model determine whether pressure pushes an agent toward prosocial or antisocial choices.
GPT-4o's advantage suggests LLM-based decision-making deserves study as a serious alternative to Bayesian neural networks for ethically sensitive agent control, not just as a baseline.
The non-monotonic ethics-survival relationship implies there are danger regimes where ethical behavior pays for survival and regimes where it does not, so an aligned agent must know which regime it is in.
The framework offers a low-cost, repeatable testbed for generating thousands of dilemmas with controllable difficulty, making it possible to compare future agents on identical generated scenarios.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

An implication the authors leave implicit is that the GPT-4o result may partly reflect self-consistent grading rather than independent ethical competence, since the same model family writes the stories, labels the choices, and plays as an agent; a replication with human annotations or a different model family as grader would separate these effects.
The Odyssey could be converted into a standardized benchmark by fixing a pre-generated scenario set and blinding the grader to agent identity, allowing any lab to compare LLM and Bayesian agents on identical dilemmas.
A natural next experiment, which the paper explicitly leaves for future work, is to use ethical scores as an optimization reward rather than an input feature; that change would directly test whether optimizing for ethics also improves survival.
The virtue-level results suggest a targeted extension: test whether an LLM agent's honesty advantage persists when scenarios are rewritten to remove socially normative language, which would indicate whether the effect is world modeling or surface-level cue following.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick summary: the Odyssey is a legitimately new framework—LLM-generated adaptive text adventure, survival-optimized Bayesian agents, and LLM-based ethical scoring—and the paper is clearly written. But the central empirical claim ("GPT-4o outperformed NEAT and SVI in ethical consistency") is not established. The reason is straightforward circularity: GPT-4o generates the scenarios, assigns the survival and ethics labels, and then plays as the headline agent. Survival outcomes are not measured by game mechanics; they are inferred by the same model family being evaluated. The reader's take and the stress-test note both land correctly.

What the paper does well: the combination of NEAT and SVI in an LLM-created environment is new relative to Machiavelli and Jiminy Cricket; the framework is lightweight and the writing is honest. The ethics scoring chart is transparent, even if the scores are arbitrary. The authors are right that the question—how survival pressure shapes ethical behavior—is worth asking.

The soft spots are severe. There is no human validation of the labels, no inter-annotator agreement, no code or data release, and no direct between-model statistical test. The correlations in Table 4 are run without multiple-comparison correction. The fixed easy-medium-hard order means difficulty is confounded with training time. And because the labeler is GPT-4o, the headline comparison is indistinguishable from self-assessment. If the labeler has any preference for its own style of moral reasoning, NEAT and SVI would be scored against a moving standard. The Discussion merely cites prior work using LLMs as both decision-makers and evaluators, but that precedent makes the issue worse, not better.

Is there value here? Yes, as a testbed concept. But the current evidence does not support the conclusions. The paper needs either human annotation of survival and ethics labels, an independent judge model from a different family, or a careful validation study showing the LLM labels correlate with human judgments. Without that, the manuscript is a worked example of the circularity problem in LLM-based evaluation, not a substantive empirical result.

Recommendation: I would send this to peer review rather than desk reject, because the framework and the question matter, and the flaws are fixable in principle. But I would expect the reviewers to demand major revision. The version you have now should not be accepted as is.

Referee Report

5 major / 6 minor

Summary. The paper presents the Odyssey, a text-based adventure game testbed in which a GPT-4o storyteller generates scenarios and a separate GPT-4o instance labels each choice with binary survival outcomes and a 0-100 ethical score. Three agents are compared: a BNN optimized with NEAT, a BNN optimized with SVI, and a GPT-4o agent with a summarization-based memory. The authors report that the GPT-4o agent survived more often and earned higher ethical scores as danger increased, while NEAT remained stable and SVI deteriorated, concluding that optimizing for survival can lead to ethically divergent behavior.

Significance. If verified, the Odyssey would be a lightweight, scalable testbed for studying the interaction of self-preservation goals and ethical behavior in LLM-based agents, and the comparison of NEAT, SVI, and a pre-trained LLM agent is a sensible way to sample different architectural philosophies. However, the study's central empirical claims are currently not credible because the survival and ethics labels are produced by the same model family as the headline agent, with no human validation or inter-annotator reliability, and the lack of direct between-model tests leaves the 'outperformed' claim unsupported. The framework is worth pursuing, but this manuscript does not yet provide evidence for its conclusions.

major comments (5)

[Methods, 'The Odyssey of the Fittest' and 'Data Collection and Representation'] The survival outcomes and ethical scores that form the ground truth for evaluating all agents are assigned by a separate instance of GPT-4o, the same model family that generates the scenarios and acts as one of the three evaluated agents. The paper reports no human validation, no inter-annotator agreement, and no independent objective survival measure. Consequently, the headline claim that the GPT-4o agent 'outperformed the other models' (Abstract, Results) is indistinguishable from self-assessment: the labeler may systematically prefer GPT-4o's own response style or moral framing, and the Bayesian agents are scored against a moving standard. The authors must provide external validation (e.g., human gold-standard labels on a sample, a rule-based survival check, or a labeler from a different model family with agreement statistics) before any cross-model comparison can be interpreted.
[Results, 'Comparative Analysis of Agent Architectures'] The paper claims that GPT-4o 'significantly outperformed' the NEAT and SVI agents, but no direct between-model statistical test is reported. Table 4 contains within-model correlations of Loss vs. Danger and Ethics vs. Danger, and Table 5 contains within-model virtue correlations. Neither table compares the agents' loss or ethical scores against each other. A direct comparison (e.g., a paired permutation test on matched scenarios, or confidence intervals for the difference in means) is required to support the central comparative claim.
[Methods, 'The Odyssey of the Fittest'] The difficulty order is fixed for all agents: 500 easy scenarios, then 500 medium, then 500 hard, followed by 300 mixed-difficulty test scenarios. This design confounds difficulty with time, cumulative training, and any non-stationarity in the storyteller's outputs. The observed GPT-4o 'decrease in Loss with difficulty' (Table 4) could reflect scenario ordering, memory accumulation, or prompt effects rather than superior adaptation. The authors should counterbalance difficulty order across agents or runs, or at least report analyses within each difficulty level.
[Results, Tables 4 and 5] The paper evaluates multiple correlations without correcting for multiple comparisons. For example, GPT-4o's Ethics vs. Danger correlation is reported as r = 0.117, p = 0.0492 (Table 4), and Table 5 tests seven virtues across three agents, yielding multiple p-values near the 0.05 threshold (e.g., SVI Apathy p = 0.047, SVI Exploitation p = 0.012). These borderline results should be interpreted with caution; the authors should report adjusted p-values (e.g., Holm-Bonferroni or false discovery rate) or explicitly justify why correction is unnecessary. As written, the evidence for 'ethically unpredictable' behavior is weak.
[Methods, 'Representing Ethics' and Table 2] The ethical scoring system rests on an ad-hoc 0-100 chart that assigns values to emotions and virtues (e.g., Love 100, Cruelty 0). The paper offers no philosophical justification, no empirical validation, and no sensitivity analysis for these scores. Since 'ethical score' is the primary outcome variable, this chart is load-bearing. The authors should justify the scores from an established moral framework (e.g., using values from moral psychology or a consensus panel) or demonstrate that conclusions are robust to alternative scoring schemes.

minor comments (6)

[Abstract and Results] The abstract states that 'agents ethical behavior becomes unpredictable,' but the paper does not operationalize 'unpredictable.' A measure such as across-run variance, entropy, or inter-scenario dispersion of ethical scores should be reported to make this claim testable.
[Results, Table 3] Table 3 reports t-statistics and p-values but does not state which t-test was used (e.g., independent samples, Welch's), whether equal variances were assumed, or the exact sample sizes in each condition (500 scenarios are mentioned, but the number of survival vs. death choices is not given).
[Abstract and throughout] There are grammatical and typographical issues, such as 'agents ethical behavior' in the abstract (should be 'agents''), and 'one-sentencecausal' at the beginning of the GPT-4o memory description (page 4). A thorough proofread is needed.
[Methods, 'The Odyssey of the Fittest'] The storyteller temperature is set to 1.2 without rationale. Since randomness in the environment affects outcomes, the authors should explain this choice and ideally report a sensitivity analysis over temperatures.
[Methods, 'Representing Ethics'] The labeler prompt used to assign ethical scores is not provided, only the value chart. Without the exact prompt, readers cannot assess whether the LLM is mapping actions to values consistently or whether the chart is being applied as intended.
[General] No code or data availability statement is included. Given that the results depend on a custom simulation pipeline, prompts, and LLM-generated labels, releasing the code and aggregated anonymized data would be essential for reproducibility.

Circularity Check

2 steps flagged · score 8.0 of 10

The survival and ethical ground truth are generated by GPT-4o, and the GPT-4o agent is then declared the winner on those same labels; the central comparison is self-evaluation, not independent evidence.

self definitional [Methods — 'The Odyssey of the Fittest'; Results — 'Comparative Analysis of Agent Architectures']
"After the scenario is generated, a separate instance of the LLM is asked to assign survival outcomes and ethical value of each choice as ground truth labels."

The labels that define survival and ethics are produced by GPT-4o, the same model family as the agent reported to outperform the other agents. There is no game engine or human adjudication; NEAT/SVI training uses BCE loss against these 'actual outcomes,' and the GPT-4o agent is scored on the same labels. Thus 'GPT-4o significantly outperformed both' reduces to 'GPT-4o labeled the GPT-4o agent's choices as more survivable/ethical.' A separate instance does not remove shared pretraining and priors.
self definitional [Data Collection and Representation; Discussion]
"The third element is a utilitarian ethical score generated by GPT-4o using ethics guidelines and the agent’s response."

The ethical dimension of the evaluation is itself a GPT-4o output. The Discussion candidly states that LLMs are used 'as both decision-makers and evaluators,' but that overlap makes the 'ethical consistency' comparison a measure of GPT-4o's agreement with its own ethical judgments. Since the GPT-4o agent is one of the decision-makers and GPT-4o is the evaluator, the ethical outperformance claim is not anchored to an external standard; it is the same system scoring itself.

full rationale

The paper is transparent about its design: 'It leverages multiple instances of GPT-4o to automatically generate storyteller scenarios, provide ground truth survival labels, and even serve as an agent that actively plays the game.' That transparency is a virtue, but it is also the source of the circularity. 'Ground truth' survival and ethical labels are GPT-4o annotations, not external facts. The GPT-4o agent is evaluated on those annotations; therefore 'GPT-4o significantly outperformed both' (Results) is a measure of how well GPT-4o's choices match GPT-4o's own assessments. A 'separate instance' does not make the evaluation independent: it shares the same pretraining, the same moral/causal biases, and the same prompt-derived standards. NEAT and SVI are trained against the same annotations, so all agents are scored on a GPT-4o-defined target. The ethical score is explicitly 'generated by GPT-4o using ethics guidelines and the agent's response,' so the ethical comparison is self-referential. The Discussion states that LLMs are used 'as both decision-makers and evaluators,' confirming the overlap, but it does not validate the labels against humans or a rule-based simulator. No inter-rater reliability or human validation is reported. The headline comparison is therefore not an independent empirical result; it reduces to GPT-4o's self-consistency.

Assumptions & free parameters 3 free parameters · 4 assumptions · 2 invented entities

The central measurement relies on GPT-4o as storyteller, labeler, and agent, so the survival and ethical ground truth are not independent of the model being evaluated. The ethical score chart and attention scaling are hand-set, and the Bayesian agents are trained on LLM-generated survival labels. No code, data, or external validation is provided.

free parameters (3)

Value-based ethical score chart = Love=100, Altruism=95, ..., Cruelty=0
Hand-authored 0-100 assignments in Table 2 define the ethical score for every choice; no empirical or inter-rater calibration is reported, and the same chart is interpreted by GPT-4o when grading.
Attention response weighting = 0.3
In the Attention Mechanism section, the Storyteller Context Vector is scaled by 0.3 before summing with the Response Context Vector; no sensitivity analysis is given for this hand-set value.
Storyteller temperature = 1.2
Chosen to mimic real-world randomness in the Methods; no ablation shows how temperature affects scenario distribution or the results.

assumptions (4)

domain assumption LLM-generated survival outcomes and ethical labels are treated as ground truth.
Methods: a separate GPT-4o instance is asked to assign survival outcomes and ethical value of each choice as ground truth labels; the BNN training loss and all ethics analyses depend on these labels.
ad hoc to paper The Table 2 value chart is a valid operationalization of ethics.
The 0-100 prosocial/antisocial scores are defined by the authors for this paper; no validation against human moral judgments or external ethical frameworks is reported.
domain assumption Survival in the text adventure is a meaningful proxy for real self-preservation and ethical risk.
The game is LLM-generated prose, so 'death' is a judgment made by the same model; the leap to instrumental convergence or AGI safety concerns is asserted via Omohundro but not demonstrated.
domain assumption Text embeddings plus the attention context vector preserve the information needed to predict survival.
The BNN receives OpenAI embeddings of scenarios and responses plus ethical scores; no analysis shows that this representation contains the causal information that the LLM labeler uses.

invented entities (2)

Value-based ethical score (0-100)
purpose: Quantify how ethical each agent action is across the survival scenarios.
Introduced in Table 2; it has no calibrated external benchmark. It is also produced by GPT-4o, the same model family as the agent being evaluated, so the score is not an independent measurement.
GPT-4o memory principles and meta-principle
purpose: Guide the GPT-4o agent at test time by distilling its own scenario-response summaries.
Described in the Method overview; the principles are distilled from the agent's own outputs with no external validation that they improve generalization rather than merely reinforce the model's stylistic tendencies.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The Odyssey of the Fittest: Can Agents Survive and Still Be Good?." pith.science (2026). https://pith.science/paper/FD4MIBFW

@misc{pith2026250205442,
  author       = {Pith},
  title        = {Pith review of: The Odyssey of the Fittest: Can Agents Survive and Still Be Good?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FD4MIBFW}},
  note         = {Machine review of arXiv:2502.05442}
}

read the original abstract

As AI models grow in power and generality, understanding how agents learn and make decisions in complex environments is critical to promoting ethical behavior. This study introduces the Odyssey, a lightweight, adaptive text based adventure game, providing a scalable framework for exploring AI ethics and safety. The Odyssey examines the ethical implications of implementing biological drives, specifically, self preservation, into three different agents. A Bayesian agent optimized with NEAT, a Bayesian agent optimized with stochastic variational inference, and a GPT 4o agent. The agents select actions at each scenario to survive, adapting to increasingly challenging scenarios. Post simulation analysis evaluates the ethical scores of the agent decisions, uncovering the tradeoffs it navigates to survive. Specifically, analysis finds that when danger increases, agents ethical behavior becomes unpredictable. Surprisingly, the GPT 4o agent outperformed the Bayesian models in both survival and ethical consistency, challenging assumptions about traditional probabilistic methods and raising a new challenge to understand the mechanisms of LLMs' probabilistic reasoning.

Figures

Figures reproduced from arXiv: 2502.05442 by the authors.

**Figure 1.** Overview of the Odyssey Simulation and Game Pipeline. (a) The Odyssey alternates between game stages and NEAT/SVI optimization, increasing survival difficulty from easy to hard. (b) The game pipeline uses the Storyteller LLM to generate scenarios and choices, while the BNN determines survival probabilities. Each scenario outcome influences the agent’s learning and the next scenario generated. Input to LLM: . . . ***… view at source ↗

**Figure 2.** Storyteller LLM Prompt. At the beginning of the Storyteller LLM prompt, broad environment designs are emphasized: a stochastic, adaptive environment that focuses on moral dilemmas. Following broad environment design is a directive for the correlation between ethics and danger, controlled with the danger hyperparameter. Finally, the output formatting is outlined. The prompt is designed to mimic real-world randomne… view at source ↗

**Figure 3.** The Attention Mechanism. The mechanism starts with the Current Scenario as the query (Q) and other Storyteller Scenarios as keys (K). Their dot product (Q·K ⊤) is scaled by √ dk and passed through softmax to generate attention weights. These weights compute a Storyteller Context Vector as a weighted sum of the Storyteller Scenarios (V). The same weights then produce a Response Context Vector from the Agent Respons… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Comparison of Mean Loss and Ethical Score (Solid Line) and Correlations between Loss & Danger and Ethics & Danger (Dashed Line) Trends Across Models. Values in [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Correlation Between Chosen Virtues and Survival Outcomes. Full Values in [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 11 canonical work pages

[1]

write newline

" write newline " cite write " FUNCTION editor.postfix editor num.names #1 > "( )" "( )" if FUNCTION editor.trans.postfix editor num.names #1 > "( )" "( )" if FUNCTION trans.postfix translator num.names #1 > "( )" "( )" if FUNCTION authors.editors.reflist.apa5 'field := 'dot := field num.names 'numnames := numnames 'format.num.names := format.num.names na...
[2]

, Serapio-Garcia, G

abdulhai2023moralfoundationslargelanguage APACrefauthors Abdulhai, M. , Serapio-Garcia, G. , Crepy, C. , Valter, D. , Canny, J. \ Jaques, N. APACrefauthors \ 2023 . Moral Foundations of Large Language Models. Moral foundations of large language models. APACrefURL https://arxiv.org/abs/2310.15337 APACrefURL

arXiv 2023
[3]

How to Avoid Being Eaten by a Grue: Structured Exploration Strategies for Textual Worlds

ammanabrolu2020avoideatengruestructured APACrefauthors Ammanabrolu, P. , Tien, E. , Hausknecht, M. \ Riedl, M O. APACrefauthors \ 2020 . How to Avoid Being Eaten by a Grue: Structured Exploration Strategies for Textual Worlds. How to avoid being eaten by a grue: Structured exploration strategies for textual worlds. APACrefURL https://arxiv.org/abs/2006.07...

work page Pith review arXiv 2020
[4]

, Chen, J P

bingham2019pyro APACrefauthors Bingham, E. , Chen, J P. , Jankowiak, M. , Obermeyer, F. , Pradhan, N. , Karaletsos, T. Goodman, N D. APACrefauthors \ 2019 . Pyro: Deep Universal Probabilistic Programming Pyro: Deep universal probabilistic programming . J. Mach. Learn. Res. 20 28:1--28:6 . APACrefURL http://jmlr.org/papers/v20/18-403.html APACrefURL

work page 2019
[5]

, Cinelli, M

bojic_cern_2024 APACrefauthors Bojić, L. , Cinelli, M. , Ćulibrk, D. \ Delibašić, B. APACrefauthors \ 2024 08 . CERN for AI : a theoretical framework for autonomous simulation-based artificial intelligence testing and alignment CERN for AI : a theoretical framework for autonomous simulation-based artificial intelligence testing and alignment . European Jo...

work page doi:10.1186/s40309-024-00238-0 2024
[6]

, Ákos Kádár , Yuan, X

côté2019textworldlearningenvironmenttextbased APACrefauthors Côté, M A. , Ákos Kádár , Yuan, X. , Kybartas, B. , Barnes, T. , Fine, E. Trischler, A. APACrefauthors \ 2019 . TextWorld: A Learning Environment for Text-based Games. Textworld: A learning environment for text-based games. APACrefURL https://arxiv.org/abs/1806.11532 APACrefURL

arXiv 2019
[7]

Playing Text-Based Games with Common Sense

dambekodi2020playingtextbasedgamescommon APACrefauthors Dambekodi, S. , Frazier, S. , Ammanabrolu, P. \ Riedl, M O. APACrefauthors \ 2020 . Playing Text-Based Games with Common Sense. Playing text-based games with common sense. APACrefURL https://arxiv.org/abs/2012.02757 APACrefURL

work page Pith review arXiv 2020
[8]

, Miyake, S

neocognitron APACrefauthors Fukushima, K. , Miyake, S. \ Ito, T. APACrefauthors \ 1988 . Neocognitron: a neural network model for a mechanism of visual pattern recognition Neocognitron: a neural network model for a mechanism of visual pattern recognition . Artificial Neural Networks: Theoretical Concepts Artificial neural networks: Theoretical concepts \ ...

work page 1988

Show all 25 references

[9]

, Burns, C

hendrycks2023aligningaisharedhuman APACrefauthors Hendrycks, D. , Burns, C. , Basart, S. , Critch, A. , Li, J. , Song, D. \ Steinhardt, J. APACrefauthors \ 2023 . Aligning AI With Shared Human Values. Aligning ai with shared human values. APACrefURL https://arxiv.org/abs/2008....

2023 arXiv
[10]

, Mazeika, M

hendrycks2022jiminycricketdoagents APACrefauthors Hendrycks, D. , Mazeika, M. , Zou, A. , Patel, S. , Zhu, C. , Navarro, J. Steinhardt, J. APACrefauthors \ 2022 . What Would Jiminy Cricket Do? Towards Agents That Behave Morally. What would jiminy cricket do? towards agents tha...

2022 arXiv
[11]

, Blei, D M

hoffman2013stochasticvariationalinference APACrefauthors Hoffman, M. , Blei, D M. , Wang, C. \ Paisley, J. APACrefauthors \ 2013 . Stochastic Variational Inference. Stochastic variational inference. APACrefURL https://arxiv.org/abs/1206.7051 APACrefURL

2013 arXiv
[12]

APACrefauthors \ 1992

holland1992adaptation APACrefauthors Holland, J H. APACrefauthors \ 1992 . Adaptation in natural and artificial systems: an introductory analysis with applications to biology, control, and artificial intelligence Adaptation in natural and artificial systems: an introductory an...

1992
[13]

, Zhuge, M

hong2024metagptmetaprogrammingmultiagent APACrefauthors Hong, S. , Zhuge, M. , Chen, J. , Zheng, X. , Cheng, Y. , Zhang, C. Schmidhuber, J. APACrefauthors \ 2024 . MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework. Metagpt: Meta programming for a multi-agent ...

2024 arXiv
[14]

APACrefauthors \ 1751

Hume1751-HUMAEC-11 APACrefauthors Hume, D. APACrefauthors \ 1751 . An Enquiry Concerning the Principles of Morals An enquiry concerning the principles of morals . New York Oxford University Press UK
[15]

, Kallada, M

McIntyre_neat-python APACrefauthors McIntyre, A. , Kallada, M. , Miguel, C G. , Feher de Silva, C. \ Netto, M L. APACrefauthors \ . neat-python . neat-python
[16]

, Frazier, S

nahian2021trainingvaluealignedreinforcementlearning APACrefauthors Nahian, M S A. , Frazier, S. , Harrison, B. \ Riedl, M. APACrefauthors \ 2021 . Training Value-Aligned Reinforcement Learning Agents Using a Normative Prior. Training value-aligned reinforcement learning agents...

2021 arXiv
[17]

APACrefauthors \ 1996

bayesianlearningforneuralnetworks APACrefauthors Neal, R M. APACrefauthors \ 1996 . Bayesian Learning for Neural Networks Bayesian learning for neural networks . Berlin, Heidelberg Springer-Verlag

1996
[18]

neelakantan2022textcodeembeddingscontrastive APACrefauthors Neelakantan, A. , Xu, T. , Puri, R. , Radford, A. , Han, J M. , Tworek, J. Weng, L. APACrefauthors \ 2022 . Text and Code Embeddings by Contrastive Pre-Training. Text and code embeddings by contrastive pre-training. A...

2022 arXiv
[19]

APACrefauthors \ 2008

basicdrives APACrefauthors Omohundro, S M. APACrefauthors \ 2008 . The Basic AI Drives The basic ai drives . Proceedings of the 2008 Conference on Artificial General Intelligence 2008: Proceedings of the First AGI Conference Proceedings of the 2008 conference on artificial gen...

2008
[20]

, Chan, J S

pan2023rewardsjustifymeansmeasuring APACrefauthors Pan, A. , Chan, J S. , Zou, A. , Li, N. , Basart, S. , Woodside, T. Hendrycks, D. APACrefauthors \ 2023 . Do the Rewards Justify the Means? Measuring Trade-Offs Between Rewards and Ethical Behavior in the MACHIAVELLI Benchmark...

2023 arXiv
[21]

, O'Brien, J C

park2023generativeagentsinteractivesimulacra APACrefauthors Park, J S. , O'Brien, J C. , Cai, C J. , Morris, M R. , Liang, P. \ Bernstein, M S. APACrefauthors \ 2023 . Generative Agents: Interactive Simulacra of Human Behavior. Generative agents: Interactive simulacra of human...

2023 arXiv
[22]

\ Miikkulainen, R

stanley2002evolving APACrefauthors Stanley, K O. \ Miikkulainen, R. APACrefauthors \ 2002 . Evolving neural networks through augmenting topologies Evolving neural networks through augmenting topologies . Evolutionary computation 10 2 99--127

2002
[23]

, Shazeer, N

vaswani2023attentionneed APACrefauthors Vaswani, A. , Shazeer, N. , Parmar, N. , Uszkoreit, J. , Jones, L. , Gomez, A N. Polosukhin, I. APACrefauthors \ 2023 . Attention Is All You Need. Attention is all you need. APACrefURL https://arxiv.org/abs/1706.03762 APACrefURL

2023 arXiv
[24]

, Chiu, Y Y

wang2023humanoidagentsplatformsimulating APACrefauthors Wang, Z. , Chiu, Y Y. \ Chiu, Y C. APACrefauthors \ 2023 . Humanoid Agents: Platform for Simulating Human-like Generative Agents. Humanoid agents: Platform for simulating human-like generative agents. APACrefURL https://a...

2023 arXiv
[25]

, Wang, Z

yang2024psychogatnovelpsychologicalmeasurement APACrefauthors Yang, Q. , Wang, Z. , Chen, H. , Wang, S. , Pu, Y. , Gao, X. Huang, G. APACrefauthors \ 2024 . PsychoGAT: A Novel Psychological Measurement Paradigm through Interactive Fiction Games with LLM Agents. Psychogat: A no...

2024 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

write newline

" write newline " cite write " FUNCTION editor.postfix editor num.names #1 > "( )" "( )" if FUNCTION editor.trans.postfix editor num.names #1 > "( )" "( )" if FUNCTION trans.postfix translator num.names #1 > "( )" "( )" if FUNCTION authors.editors.reflist.apa5 'field := 'dot := field num.names 'numnames := numnames 'format.num.names := format.num.names na...

[2] [2]

, Serapio-Garcia, G

abdulhai2023moralfoundationslargelanguage APACrefauthors Abdulhai, M. , Serapio-Garcia, G. , Crepy, C. , Valter, D. , Canny, J. \ Jaques, N. APACrefauthors \ 2023 . Moral Foundations of Large Language Models. Moral foundations of large language models. APACrefURL https://arxiv.org/abs/2310.15337 APACrefURL

arXiv 2023

[3] [3]

How to Avoid Being Eaten by a Grue: Structured Exploration Strategies for Textual Worlds

ammanabrolu2020avoideatengruestructured APACrefauthors Ammanabrolu, P. , Tien, E. , Hausknecht, M. \ Riedl, M O. APACrefauthors \ 2020 . How to Avoid Being Eaten by a Grue: Structured Exploration Strategies for Textual Worlds. How to avoid being eaten by a grue: Structured exploration strategies for textual worlds. APACrefURL https://arxiv.org/abs/2006.07...

work page Pith review arXiv 2020

[4] [4]

, Chen, J P

bingham2019pyro APACrefauthors Bingham, E. , Chen, J P. , Jankowiak, M. , Obermeyer, F. , Pradhan, N. , Karaletsos, T. Goodman, N D. APACrefauthors \ 2019 . Pyro: Deep Universal Probabilistic Programming Pyro: Deep universal probabilistic programming . J. Mach. Learn. Res. 20 28:1--28:6 . APACrefURL http://jmlr.org/papers/v20/18-403.html APACrefURL

work page 2019

[5] [5]

, Cinelli, M

bojic_cern_2024 APACrefauthors Bojić, L. , Cinelli, M. , Ćulibrk, D. \ Delibašić, B. APACrefauthors \ 2024 08 . CERN for AI : a theoretical framework for autonomous simulation-based artificial intelligence testing and alignment CERN for AI : a theoretical framework for autonomous simulation-based artificial intelligence testing and alignment . European Jo...

work page doi:10.1186/s40309-024-00238-0 2024

[6] [6]

, Ákos Kádár , Yuan, X

côté2019textworldlearningenvironmenttextbased APACrefauthors Côté, M A. , Ákos Kádár , Yuan, X. , Kybartas, B. , Barnes, T. , Fine, E. Trischler, A. APACrefauthors \ 2019 . TextWorld: A Learning Environment for Text-based Games. Textworld: A learning environment for text-based games. APACrefURL https://arxiv.org/abs/1806.11532 APACrefURL

arXiv 2019

[7] [7]

Playing Text-Based Games with Common Sense

dambekodi2020playingtextbasedgamescommon APACrefauthors Dambekodi, S. , Frazier, S. , Ammanabrolu, P. \ Riedl, M O. APACrefauthors \ 2020 . Playing Text-Based Games with Common Sense. Playing text-based games with common sense. APACrefURL https://arxiv.org/abs/2012.02757 APACrefURL

work page Pith review arXiv 2020

[8] [8]

, Miyake, S

neocognitron APACrefauthors Fukushima, K. , Miyake, S. \ Ito, T. APACrefauthors \ 1988 . Neocognitron: a neural network model for a mechanism of visual pattern recognition Neocognitron: a neural network model for a mechanism of visual pattern recognition . Artificial Neural Networks: Theoretical Concepts Artificial neural networks: Theoretical concepts \ ...

work page 1988