Pith. sign in

REVIEW 2 major objections 57 references

Are Humans Evolved Instruction Followers? An Underlying Inductive Bias Enables Rapid Instructed Task Learning

T0 review · 2 major / 0 minor · reviewed 2026-06-30 · grok-4.3

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

Pith's one-line read Humans possess an evolved instruction-following bias that enables rapid instructed task learning from language alone.

desk verdict This is a position paper hypothesizing an evolved human instruction-following bias analogous to LLM tuning, but it rests on synthesis without new data or tests to back the innate claim. read the letter →

arxiv 2606.29792 v1 pith:RPTSETP4 submitted 2026-06-29 cs.CL

Anjishnu Kumar This is my paper · ORCID
classification cs.CL
keywords instruction-followingbiasrapidinstructedtasklearninginductivehumancognitionlargelanguagemodelszero-shotperformancecognitiveflexibilitypositionpaper
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

The paper argues that an inductive bias shaped by evolution lets humans interpret and carry out novel verbal or written instructions correctly on the first try. This bias is presented as the key mechanism behind rapid instructed task learning, a form of cognitive flexibility that current artificial systems achieve only through explicit training. The authors draw an explicit parallel to how large language models reach zero-shot performance after instruction tuning. Evidence is synthesized from cognitive science, neuroscience, and machine learning to support the claim that this capacity is innate in humans rather than learned. The position calls for interdisciplinary tests of the idea as a unifying account of fast generalization in both natural and artificial networks.

What carries the argument

The evolved instruction-following bias, an innate cognitive architecture feature that allows interpretation and execution of linguistic instructions without prior task-specific training.

What would settle it

Demonstration that instruction-following performance on novel tasks emerges at the same rate and reliability in populations or species lacking any evolutionary history of linguistic instruction, or that it can be fully replicated in artificial systems without any analogue to the proposed bias.

Watch

Extended reading notes

Core claim

Humans possess an evolved instruction-following bias—an inductive bias shaped by evolution to interpret and execute linguistic instructions—which critically enables fast generalization of behavior from language and functions analogously to the way large language models leverage instruction tuning to achieve zero-shot task performance.

Load-bearing premise

Instruction-following arises as an innate evolved feature of cognitive architecture rather than through learning or cultural transmission alone.

Editorial extensions

If this is right

  • Instruction-following in artificial systems could be treated as an architectural prior rather than solely a product of specialized training protocols.
  • Testable predictions can be derived about the developmental trajectory and neural signatures of instruction interpretation in humans.
  • Rapid instructed task learning would depend on this pre-existing bias rather than general learning mechanisms alone.
  • Interdisciplinary work could treat instruction-following as a shared mechanism across natural and artificial neural networks.

Reading between the lines

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

  • The bias might account for why humans can generalize from instructions across domains where current models still require extensive fine-tuning.
  • It would imply that language evolution was shaped in part by selection pressures favoring instruction-based coordination.
  • Disorders affecting instruction comprehension could be re-examined as disruptions of this specific inductive bias rather than general executive function deficits.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 0 minor

Summary. The manuscript is a position paper claiming that rapid instructed task learning (RITL) in humans is enabled by an evolved 'instruction-following bias'—an innate inductive bias shaped by evolution to interpret and execute linguistic instructions. This bias is posited to function analogously to instruction tuning in LLMs, arising as a feature of cognitive architecture rather than through learning or cultural transmission. The paper synthesizes existing findings from cognitive science, neuroscience, and machine learning, outlines testable predictions, and calls for interdisciplinary research on instruction-following as a unifying mechanism.

Significance. If the hypothesis holds, it could provide a cross-disciplinary framework linking human cognitive flexibility to mechanisms in artificial neural networks, potentially guiding new empirical tests and AI model designs. The manuscript's strengths include its explicit synthesis of multi-field evidence and its formulation of falsifiable predictions, which creditably moves beyond pure speculation even without new data or derivations.

major comments (2)
  1. [Abstract] Abstract: The central claim that the instruction-following bias 'arises as an innate cognitive architecture feature' (rather than primarily through learning or cultural transmission) is load-bearing for the evolutionary hypothesis but rests solely on literature synthesis without new comparative evidence or a formal model that would falsify learning-based alternatives.
  2. [Abstract] Abstract and hypothesis framing: The LLM analogy is presented as key support, yet instruction tuning's success is defined by the same zero-shot instructed performance the paper seeks to explain in humans; this creates a circular mapping without independent grounding or distinguishing data.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments on our position paper. We address each major comment below, clarifying the scope of our hypothesis while noting the inherent limitations of a synthesis-based approach without new data.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that the instruction-following bias 'arises as an innate cognitive architecture feature' (rather than primarily through learning or cultural transmission) is load-bearing for the evolutionary hypothesis but rests solely on literature synthesis without new comparative evidence or a formal model that would falsify learning-based alternatives.

    Authors: We acknowledge that the claim is presented as a hypothesis grounded in cross-disciplinary literature synthesis rather than new empirical evidence or a formal model capable of directly falsifying learning-based accounts. As a position paper, the contribution lies in formulating the evolutionary perspective and deriving testable predictions for future work that could adjudicate between innate bias and acquired alternatives. We will revise the abstract and introduction to more explicitly label the claim as a hypothesis and to highlight the absence of new comparative data as a limitation. revision: partial

  2. Referee: [Abstract] Abstract and hypothesis framing: The LLM analogy is presented as key support, yet instruction tuning's success is defined by the same zero-shot instructed performance the paper seeks to explain in humans; this creates a circular mapping without independent grounding or distinguishing data.

    Authors: The analogy is offered as a mechanistic parallel rather than direct empirical support: both systems exhibit rapid instructed task learning once an instruction-following inductive bias is in place, with the bias acquired through tuning in LLMs and posited as evolved in humans. We agree that the shared performance metric alone does not distinguish the accounts and will revise the relevant sections to separate the observable RITL behavior from the proposed underlying bias mechanisms, including differences in acquisition. We note that fully independent grounding would require targeted experiments beyond the scope of this position paper. revision: partial

Circularity Check

0 steps flagged · score 0.0 of 10

Position paper posits hypothesis via synthesis and analogy; no derivation reduces to inputs

full rationale

This is a position paper that synthesizes existing findings from cognitive science, neuroscience, and ML, then explicitly posits ('we posit') that instruction-following is an innate evolved bias in humans, functioning analogously to LLM instruction tuning. No equations, parameter fits, or derivations appear in the provided text. The central claim is framed as a hypothesis to be tested, not a result obtained by construction from the LLM parallel or from self-citations. The analogy is presented as functional similarity rather than definitional equivalence or fitted prediction. No load-bearing self-citation chains or ansatzes smuggled via prior work are evident. The paper remains self-contained as an argument for future interdisciplinary investigation rather than a closed derivation.

Assumptions & free parameters 0 free parameters · 2 assumptions · 1 invented entities

The claim rests on the existence of an evolved innate bias whose only support in the provided text is analogy to current AI training practices and prior cognitive findings; no independent evidence or falsifiable handle is supplied.

assumptions (2)
  • domain assumption Human adults can perform novel tasks correctly on the first attempt after receiving verbal or written instructions.
    Stated as the starting observation in the abstract.
  • ad hoc to paper Instruction-following in humans arises as an innate cognitive architecture feature shaped by evolution.
    Central hypothesis presented without derivation or new data.
invented entities (1)
  • evolved instruction-following bias
    purpose: To explain rapid instructed task learning as an innate feature rather than learned behavior.
    Postulated as the load-bearing mechanism; no independent evidence such as a predicted neural signature or genetic marker is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Are Humans Evolved Instruction Followers? An Underlying Inductive Bias Enables Rapid Instructed Task Learning." pith.science (2026). https://pith.science/paper/RPTSETP4

@misc{pith2026260629792,
  author       = {Pith},
  title        = {Pith review of: Are Humans Evolved Instruction Followers? An Underlying Inductive Bias Enables Rapid Instructed Task Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RPTSETP4}},
  note         = {Machine review of arXiv:2606.29792}
}
read the original abstract

Human adults can often perform a novel task correctly on the first attempt after only receiving verbal or written instructions. This rapid instructed task learning (RITL) is a hallmark of human cognitive flexibility, yet its mechanisms and parallels in artificial systems remain under-explored across disciplines. In this position paper, we argue that humans possess an evolved instruction-following bias -- an inductive bias shaped by evolution to interpret and execute linguistic instructions which critically enables fast generalization of behavior from language. This bias functions analogously to the way large language models (LLMs) leverage instruction tuning to achieve zero-shot task performance. We synthesize evidence from cognitive science, neuroscience, and machine learning research to support this hypothesis. While instruction-following in AI is currently achieved via specialized training protocols, we posit that in humans it arises as an innate cognitive architecture feature. We outline testable predictions and call for more interdisciplinary research to investigate Instruction-Following as a unifying mechanism enabling rapid task learning in both natural and artificial neural networks.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 3 canonical work pages

  1. [1]

    Frontiers in Human Neuroscience , volume=

    Rapid instructed task learning: A new window into the human brain's unique capacity for cognitive control , author=. Frontiers in Human Neuroscience , volume=. 2013 , doi=

    2013

  2. [2]

    Trends in Cognitive Sciences , volume=

    The multiple-demand (MD) system of the primate brain: mental programs for intelligent behaviour , author=. Trends in Cognitive Sciences , volume=. 2010 , doi=

    2010

  3. [3]

    Proceedings of the National Academy of Sciences , volume=

    Multi-task connectivity reveals flexible hubs for adaptive task control , author=. Proceedings of the National Academy of Sciences , volume=. 2013 , doi=

    2013

  4. [4]

    Frontiers in Human Neuroscience , volume=

    Rapid transfer of abstract rules to novel contexts in human lateral prefrontal cortex , author=. Frontiers in Human Neuroscience , volume=. 2011 , doi=

    2011

  5. [5]

    Annual Review of Neuroscience , volume=

    An integrative theory of prefrontal cortex function , author=. Annual Review of Neuroscience , volume=. 2001 , doi=

    2001

  6. [6]

    Philosophical Transactions of the Royal Society B , volume=

    Natural pedagogy as evolutionary adaptation , author=. Philosophical Transactions of the Royal Society B , volume=. 2011 , doi=

    2011

  7. [7]

    Trends in Cognitive Sciences , volume=

    Natural pedagogy , author=. Trends in Cognitive Sciences , volume=. 2009 , doi=

    2009

  8. [8]

    Frontiers in Psychology , volume=

    Frontostriatal mechanisms in instruction-based learning , author=. Frontiers in Psychology , volume=. 2012 , doi=

    2012

Show all 57 references
  1. [9]

    Frontiers in Human Neuroscience , volume=

    When planning results in loss of control: intention-based reflexivity and working memory , author=. Frontiers in Human Neuroscience , volume=. 2012 , doi=

    2012

  2. [10]

    Neuron , volume=

    Neural activity in the primate prefrontal cortex during associative learning , author=. Neuron , volume=. 1998 , doi=

    1998

  3. [11]

    1993 , publisher=

    Language Comprehension in Ape and Child , author=. 1993 , publisher=

    1993

  4. [12]

    American Journal of Physical Anthropology , volume=

    Prefrontal cortex in humans and apes: A comparative study of area 10 , author=. American Journal of Physical Anthropology , volume=. 2001 , doi=

    2001

  5. [13]

    Science , volume=

    Why copy others? Insights from the social learning strategies tournament , author=. Science , volume=. 2010 , doi=

    2010

  6. [14]

    Nature Communications , volume=

    Alignment of brain embeddings and artificial contextual embeddings in natural language points to common geometric patterns , author=. Nature Communications , volume=. 2024 , doi=

    2024

  7. [15]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Language Models are Few-Shot Learners , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  8. [16]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Training language models to follow instructions with human feedback , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  9. [17]

    International Conference on Learning Representations (ICLR) , year=

    Finetuned Language Models Are Zero-Shot Learners , author=. International Conference on Learning Representations (ICLR) , year=

  10. [18]

    arXiv preprint arXiv:2210.11416 , year=

    Scaling Instruction-Finetuned Language Models , author=. arXiv preprint arXiv:2210.11416 , year=

    arXiv

  11. [19]

    International Conference on Learning Representations (ICLR) , year=

    Multitask Prompted Training Enables Zero-Shot Task Generalization , author=. International Conference on Learning Representations (ICLR) , year=

  12. [20]

    2018 , url=

    Reinforcement Learning: An Introduction (2nd ed.) , author=. 2018 , url=

    2018

  13. [21]

    and Gergely, G

    Csibra, G. and Gergely, G. , title =. Trends in Cognitive Sciences , year =

  14. [22]

    and Gergely, G

    Csibra, G. and Gergely, G. , title =. Philosophical Transactions of the Royal Society B , year =

  15. [23]

    Lyons, D. E. and Young, A. G. and Keil, F. C. , title =. Proceedings of the National Academy of Sciences , year =

  16. [24]

    and Tomaselli, K

    Nielsen, M. and Tomaselli, K. , title =. Psychological Science , year =

  17. [25]

    Longman, C. S. and colleagues , title =. Journal of Cognition , year =. doi:10.5334/joc.63 , url =

    doi

  18. [26]

    and Cole, M

    Meiran, N. and Cole, M. W. and Braver, T. S. , title =. Frontiers in Human Neuroscience , year =

  19. [27]

    and Pereg, M

    Meiran, N. and Pereg, M. and Givon, E. and Danieli, G. and Shahar, N. , title =. Neuropsychologia , year =

  20. [28]

    and Wolfensteller, U

    Ruge, H. and Wolfensteller, U. , title =. Cerebral Cortex , year =

  21. [29]

    and Waszak, F

    Hartstra, E. and Waszak, F. and colleagues , title =. Human Brain Mapping , year =

  22. [30]

    Cole, M. W. and Reynolds, J. R. and Power, J. D. and Repovs, G. and Anticevic, A. and Braver, T. S. , title =. Nature Neuroscience , year =

  23. [31]

    Waskom, M. L. and Kumaran, D. and Gordon, A. M. and Rissman, J. and Wagner, A. D. , title =. Neuron , year =

  24. [32]

    Verbal instructions induce persistent neural activity in the prefrontal cortex , journal =

    Gonz. Verbal instructions induce persistent neural activity in the prefrontal cortex , journal =. 2017 , volume =

    2017

  25. [33]

    , title =

    Duncan, J. , title =. Trends in Cognitive Sciences , year =

  26. [34]

    and Armstrong, E

    Semendeferi, K. and Armstrong, E. and Schleicher, A. and Zilles, K. and Van Hoesen, G. W. , title =. American Journal of Physical Anthropology , year =

  27. [35]

    Savage-Rumbaugh, E. S. and Murphy, J. and Sevcik, R. A. and Brakke, K. E. and Williams, S. L. and Rumbaugh, D. M. and Bates, E. , title =. Monographs of the Society for Research in Child Development , year =

  28. [36]

    and Call, J

    Kaminski, J. and Call, J. and Fischer, J. , title =. Science , year =

  29. [37]

    Pilley, J. W. and Reid, A. K. , title =. Behavioural Processes , year =

  30. [38]

    Pepperberg, I. M. , title =. Applied Animal Behaviour Science , year =

  31. [39]

    and Bosma, M

    Wei, J. and Bosma, M. and Zhao, V. and others , title =. NeurIPS , year =

  32. [40]

    and Webson, A

    Sanh, V. and Webson, A. and Raffel, C. and others , title =. ICLR , year =

  33. [41]

    and Wu, J

    Ouyang, L. and Wu, J. and Jiang, X. and others , title =. NeurIPS , year =

  34. [42]

    Chung, H. W. and Hou, L. and Longpre, S. and others , title =. JMLR , year =

  35. [43]

    and Kordi, Y

    Wang, Y. and Kordi, Y. and Mishra, S. and others , title =. arXiv , year =

  36. [44]

    Aw, K. L. and Montariol, S. and AlKhamissi, B. and Schrimpf, M. and Bosselut, A. , title =. Conference on Language Modeling (COLM) , year =

  37. [45]

    and colleagues , title =

    Gao, C. and colleagues , title =. bioRxiv , year =

  38. [46]

    Lake, B. M. and Baroni, M. , title =. ICML , year =

  39. [47]

    Keysers, D. and Sch. Measuring compositional generalization: A comprehensive method on realistic data , booktitle =. 2020 , url =

    2020

  40. [48]

    and Brohan, A

    Ahn, M. and Brohan, A. and Brown, N. and others , title =. arXiv , year =

  41. [49]

    and Brown, N

    Brohan, A. and Brown, N. and Carbajal, J. and others , title =. arXiv , year =

  42. [50]

    and Gupta, A

    Jiang, Y. and Gupta, A. and Zhang, Z. and others , title =. NeurIPS , year =

  43. [51]

    , title =

    Truswell, R. , title =. 2017 , url =

    2017

  44. [52]

    Animal cognition , volume=

    Causal knowledge and imitation/emulation switching in chimpanzees (Pan troglodytes) and children (Homo sapiens) , author=. Animal cognition , volume=. 2005 , publisher=

    2005

  45. [53]

    modelling on imitation and overimitation , author=

    Effects of verbal instruction vs. modelling on imitation and overimitation , author=. Humanities and Social Sciences Communications , volume=. 2021 , publisher=

    2021

  46. [54]

    and Brown, M

    Hare, B. and Brown, M. and Williamson, C. and Tomasello, M. , title =. Science , year =

  47. [55]

    Kirchhofer, K. C. and Zimmermann, F. and Kaminski, J. and Tomasello, M. , title =. PLOS ONE , year =

  48. [56]

    Animal cognition , volume=

    Comprehension of human pointing gestures in young human-reared wolves (Canis lupus) and dogs (Canis familiaris) , author=. Animal cognition , volume=. 2008 , publisher=

    2008

  49. [57]

    arXiv preprint arXiv:2312.00575 , year=

    Instruction-tuning aligns llms to the human brain , author=. arXiv preprint arXiv:2312.00575 , year=

Pith tools

Reviewed June 30, 2026 · model on record in the stance chip above.