Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

In an inspectable agent, a recurrent persistence loop and an affect proxy produce two cleanly separable families of consciousness-indicator-like signatures — persistence on one side, preference stability, scanning, and caution on the other.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 20:25 UTC pith:U3BQYCJT

load-bearing objection A transparent, reproducible toy demonstration that indicator-like signatures can be engineered; the component attribution is partly built into the design, but that is the point. the 4 major comments →

arxiv 2602.23232 v2 pith:U3BQYCJT submitted 2026-02-26 cs.AI

ReCoN-Ipsundrum: An Inspectable Recurrent Persistence Loop Agent with Affect-Coupled Control and Mechanism-Linked Consciousness Indicator Assays

Aishik Sanyal This is my paper
classification cs.AI
keywords machine consciousness indicatorsipsundrumrecurrent processingaffect couplingqualiaphiliaexploratory playcausal lesionReCoN
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper claims that in a small, fully inspectable agent, two distinct mechanisms produce two cleanly separated families of consciousness-indicator-like signatures. A recurrent 'ipsundrum' loop that sustains sensory salience after a stimulus is removed causally produces post-stimulus persistence: lesioning it drops persistence by about 20–28%, while a baseline without recurrence shows no drop. Adding an affect proxy (valence, arousal, body-budget) that enters the action score produces stable scenic-route preference under novelty competition, structured local scanning in exploratory play, and lingering planned caution after a pain-like contact — none of which appear with recurrence alone. The larger point is that indicator-like behaviors can be engineered by design choices, so behavioral markers alone are not enough to attribute anything to a mechanism; architectural inspection and causal intervention are required.

Core claim

Using three fixed-parameter variants — a ReCoN baseline, the baseline plus a recurrent ipsundrum loop over sensory salience, and the loop plus an affect proxy — the paper reports a novelty dissociation: non-affect variants shift their scenic-route choice toward the more novel lane (Δscenic-entry = 0.07), while the affect variant stays stable (Δscenic-entry = 0.01) even when the scenic lane is the less novel one (median Δnovelty ≈ −0.43). In reward-free play, only the affect variant shows structured local investigation (31.4 scan events vs 0.9; cycle score 7.6). In a pain-tail probe, only the affect variant sustains prolonged planned caution (turn-rate tail duration 90 vs 5). A within-episode

What carries the argument

The machinery is (1) the ipsundrum recurrence: a single-step state update that mixes current sensory drive with a low-pass-filtered efference-copy signal (N_e) and a 'thick-moment' integrator, so that a transient stimulus leaves a persisting trace in the salience variable N_s; and (2) the optional affect proxy, a body-budget model with valence/arousal readouts that can modulate the loop's precision/gain and enters the action-scoring equation alongside salience, epistemic curiosity, novelty, goal progress, and hazard terms. The causal-lesion assay zeroes the feedback and integration flags at t=3 to attribute persistence directly to the recurrence.

Load-bearing premise

The load-bearing premise is that the difference between the Ipsundrum and Ipsundrum+affect variants isolates the effect of affect coupling, but several design changes are made at once — negative drive is rectified away in the non-affect variant, affect terms are added to the action score, and a separate arousal-gated caution penalty is introduced — and the lesion result is partly circular because persistence is implemented by the very recurrence that is removed.

What would settle it

Run the same suite with four additional fixed-parameter variants that independently toggle: (a) signed vs rectified sensory drive with no affect terms, (b) affect terms in the score with no arousal-gated forward penalty, (c) arousal-gated forward penalty with no affect terms, and (d) a non-recurrent forward model with signed drive and affect terms. If the 90-step caution tail appears in (c) alone, or if the stable scenic preference disappears in (b), the paper's affect-coupling attribution is falsified; if both require affect terms plus the penalty, the dissociation is specific. A second falsi

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the dissociations hold, recurrence and affect coupling are independently necessary for the respective indicator-like signatures: recurrence alone gives persistence, but persistence does not yield scanning or stable preference; affect coupling is what converts persistence into planned caution and structured exploration.
  • The lesion result implies that the persistence marker can be causally attributed to the implemented feedback/integration mechanism, and that the baseline script+planning substrate contributes nothing to that signature.
  • The corridor result implies that a preference that looks like 'qualiaphilia' (sensory experience for its own sake) can actually be value-shaped — the scenic lane's negative (beneficial) input changes internal valence/arousal and thereby the action score — so the behavior does not by itself indicate a hedonic preference independent of design.
  • Treating indicators as credence-shifting only when paired with mechanistic hypotheses: the paper's broader methodological claim is that any single behavioral marker is gameable and should be accompanied by architectural inspection and causal interventions.
  • The dissociation table (persistence with recurrence; stable preference, scanning, caution with affect) provides a concrete template for how to attribute behavior to components in other minimal agents.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • An implicit consequence of the corridor result: 'reward-free' is not the same as 'incentive-free' — the scenic lane's benign sensory input is deposited into the body budget and valence terms, so an affect-coupled agent can show stable 'liking' without any external reward; this cautions against treating seemingly intrinsic preferences in minimal agents as evidence of qualia-like value.
  • A natural next step the paper leaves implicit: independently lesion the affect weights, the signed drive, and the arousal-gated forward penalty in the affect variant. If removing only the affect weights eliminates the 90-step caution tail and the scan events, the coupling story is strengthened; if the caution persists without them, the hand-set 'arousal-gated caution penalty' is the true carrier o
  • The paper's own transparency about value-shaping points to a lesson it states only implicitly: indicator-based consciousness assessment should routinely control for the experimenter's hand in wiring preferences into the agent, much as psychophysics controls for demand characteristics.
  • The lesion-and-dissociation protocol could be carried over to recurrent neural networks: one could test whether measured persistence and behavioral caution co-vary with genuine recurrent path dependence rather than with input rectification or hand-set penalties.

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

4 major / 5 minor

Summary. The paper presents ReCoN-Ipsundrum, a small, inspectable agent built on a ReCoN state machine, augmented with a recurrent 'ipsundrum' persistence loop over sensory salience N_s and an optional constructionist affect/interoception proxy (valence/arousal/body budget). Three fixed-parameter variants are compared (ReCoN, Ipsundrum, Ipsundrum+affect) across four assays: goal-directed navigation, a familiarity-controlled corridor preference task (qualiaphilia), reward-free exploratory play, a pain-tail probe, and a within-episode causal lesion of recurrence. The headline claims are that recurrence causally supports post-stimulus persistence in N_s, and that affect-coupled control produces valence-stable scenic preference, structured local scanning, and lingering planned caution. The paper explicitly disclaims any consciousness attribution and positions the work as a demonstration of how indicator-like signatures can be engineered and why mechanistic/causal evidence is needed.

Significance. If the dissociations were cleanly established, the paper would be a useful, sobering demonstration that theory-inspired indicators can be manufactured by small architectural choices, and that behavioral markers alone are insufficient. The paper's strengths include: full released code and artifacts, deterministic seeds and bootstrap CIs, a transparent statement of design choices, self-disclosed limitations (e.g., the circularity of the lesion check), and a clear separation between 'indicator', 'marker', and 'claim'. The large effect sizes in scanning (31.4 vs. 0.9), tail duration (90 vs. 5), and GridWorld success (0.99 vs. 0.50) are striking. However, the central attribution of the affect-associated signatures to 'affect coupling' is confounded by simultaneous design changes, and the key novelty-sensitivity dissociation has a confidence interval including zero. The paper is therefore valuable as a cautionary engineering case study, but the mechanistic attribution claims go beyond what the current ablation design supports.

major comments (4)
  1. [Model Summary, 'Sensory Drive and Terminal Semantics'; Eq. (8)] The affect-vs-non-affect comparison does not isolate 'affect coupling'. Three changes vary simultaneously: (i) non-affect variants rectify the sensory drive (I_t <- max(0, I_t)), discarding negative/scenic input, while the affect variant processes signed I_t and deposits pleasant input into the budget; (ii) Eq. (8) gains affect terms w_v N_v + w_a N_a + w_bb|bb-sp| with hand-set weights; (iii) an explicit 'arousal-gated caution penalty for forward moves' is added in the Policy section. Corridor stability is value-shaped by (i)-(ii), and the pain-tail lingering caution is directly implemented by (iii). With no ablation varying these components independently, the conclusion that 'affect-coupled control → preference stability, scanning, and lingering caution' is not established beyond the design that created it. This is the load-bearing attribution of the paper and needs either component-wi
  2. [Qualiaphilia (Familiarity-Controlled Novelty Competition), Figure 1] The novelty dissociation is not statistically supported as reported. The non-affect variants show Δscenic-entry = 0.07 with 95% CI [-0.02, 0.16], which includes zero; the affect variant shows 0.01 [0.00, 0.03]. The two CIs overlap substantially. The paper's claim that 'non-affect variants are novelty-sensitive' therefore rests on a point estimate whose interval includes the null. To support a dissociation, the report should provide a contrast (difference in Δscenic-entry) with its CI, or a test of interaction. The current presentation overstates the evidence for the very dissociation that motivates the paper.
  3. [Causal Lesion and 'Addressing circularity in the lesion result'] The lesion result is admitted to be circular: lesioning recurrence reduces persistence partly because persistence is implemented by recurrence. The paper reframes the lesion as an 'implementation-fidelity and causal-attribution check', but the Conclusion still states 'recurrence → post-stimulus persistence' as a component attribution. That attribution is not a discovery about the mechanism; it is a restatement of the implementation. The dissociation between persistence and scanning/caution is more informative, but the lesion itself cannot independently support the persistence claim. Please either remove the causal language from the Conclusion or add a control implementation where persistence is implemented differently (e.g., by a non-recurrent time-integration buffer) and show that the lesion does not affect that signature.
  4. [Policy section and Pain-Tail assay] The 'arousal-gated caution penalty for forward moves' is described only as an 'additional small term in code' and is not included in Eq. (8), yet it is the most plausible direct cause of the prolonged planned caution in the pain-tail assay (tail duration 90 vs. 5). Because this term is entangled with the affect manipulation and is not reported with a parameter value or ablated separately, the claim that 'only Ipsundrum+affect shows prolonged planned caution' (Table 5) is not a clean test of affect coupling; it may simply reflect the presence of the explicit caution penalty. This component should be surfaced in the main equations, given a parameter value, and ablated independently to establish whether it alone produces the tail-duration effect.
minor comments (5)
  1. [Evaluated Variants] The text says the variants 'differ only in internal dynamics and which internal variables exist', but the non-affect variants also rectify the sensory drive I_t. This is a difference in input processing, not merely internal dynamics, and should be stated explicitly in the variant description.
  2. [Table 5] The first column header reads 'Recon' while the text uses 'ReCoN'. Please standardize. Also, the ✓/× symbols should be defined in the caption.
  3. [Pain-Tail assay] The phrase 'post-stimulus N_s AUC above baseline: Ipsundrum≈0.24; Ipsundrum+affect≈0.15' is ambiguous: AUC is a time-integrated quantity, so the units should be stated (e.g., 'sensor-units × steps'). Also, the text says 'half-life collapses to 0 for all variants' but then uses AUC; this should be explained more clearly.
  4. [Model Summary, Eq. (7)] The parameter 'h' in Eq. (7) is defined only through Eq. (4) ('h X_t') but is not given a default value in the text. Please add a table of all free parameters and their default values, ideally in one place, since the fixed-parameter design is central to the paper.
  5. [Discussion, 'Estimation uncertainty'] The paper says 'some intervals remain wide (especially when metrics are coarse or censored)', but the headline GridWorld success CI '0.99[0.97,1.00]' is extremely narrow. A more systematic treatment of interval width across all main readouts would help readers calibrate confidence in the dissociations.

Circularity Check

3 steps flagged

Three of the four headline dissociations are installed by the architecture: the lesion persistence result restates the recurrence equations, the corridor stability is value-shaped by signed sensory drive plus affect weights, and the pain-tail caution is the explicitly added arousal-gated penalty.

specific steps
  1. self definitional [Discussion, 'Addressing circularity in the lesion result'; Eqs. (1)-(6); 'Causal Lesion (Attributing Persistence to Recurrence/Integration)']
    "A fair critique is that lesioning recurrence reduces persistence partly because persistence is implemented by recurrence. We treat the lesion primarily as an implementation-fidelity and causal-attribution check: it establishes that the persistence signature is not an incidental artifact."

    The persistence signature (post-stimulus N^s AUC) is produced by the very recurrence/integration terms the lesion disables: E_{t-1}, pi_t, and d in Eqs. (1)-(6). ReCoN has no such terms. Reporting an AUC drop after deleting these terms is a consistency check, not an independent discovery of a 'recurrence -> persistence' link; the paper itself calls the critique fair, yet the Conclusion still lists this as a clean component attribution.

  2. self definitional [Policy: Short-Horizon Internal Rollout with Model-Aligned Forward Dynamics; 'Pain-Tail' Results; Conclusion]
    "Additional small terms in code implement a low-change epistemic penalty, a mild forward prior, an arousal-gated caution penalty for forward moves (to link high arousal to avoidance), and a small hazard-touch penalty proportional to the predicted touch sensor. ... Only Ipsundrum+affect shows prolonged planned 'caution' (turn-rate tail duration≈90[52, 128] vs. 5 and 5)."

    The measured 'lingering planned caution' is directly caused by the hand-added 'arousal-gated caution penalty for forward moves,' whose stated purpose is 'to link high arousal to avoidance.' After the forced hazard, elevated arousal N_a makes forward moves score lower, so the agent keeps turning for ~90 steps. Non-affect variants have no N_a and no such term, hence tail duration 5. The claimed affect-coupled-control finding is therefore the explicit policy term by another name.

  3. self definitional [Model Summary, 'Sensory Drive and Terminal Semantics'; 'Qualiaphilia' Results; Discussion, 'Construct validity of interoception and qualiaphilia']
    "when affect is disabled we rectify negative input (I_t ← max(0, I_t)), so non-affect variants can represent cost but do not obtain a built-in 'pleasantness' benefit from negative I_t. ... Because scenic vs. dull changes the signed sensory drive I_t, this stability is value-shaped in our implementation ... so 'stable scenic preference' here reflects affect coupling rather than value-neutral sensory richness."

    The stable scenic preference is installed by construction: the affect variant processes signed I_t, and Eq. (8) gives positive weight to valence and body-budget terms, so scenic (negative I_t) improves the action score; non-affect variants rectify I_t and lack the affect scoring terms. The 'affect coupling -> preference stability' dissociation is therefore a comparison between an agent with an in-built hedonic bias for scenic input and agents without it, not an emergent behavioral discovery. The paper concedes the result is 'value-shaped.'

full rationale

The paper is unusually transparent, explicitly labeling the lesion critique as fair and the corridor stability as value-shaped. Those admissions make the circularity visible rather than hidden. The lesion result is an implementation check on Eqs. (1)-(6): recurrence is what the integrator/feedback terms do, so lesioning them and observing lower N^s AUC restates the constructor. The pain-tail caution is the direct output of an 'arousal-gated caution penalty for forward moves' whose explicit purpose matches the measured tail duration. The corridor-preference stability follows from signed sensory drive plus the affect weights in Eq. (8), while non-affect variants rectify the same signal away; the affect manipulation is thus bundled with reward shaping and rectification. These are by-construction effects, not independent first-principles predictions. Scanning is the one headline signature not transparently hard-wired, and no self-citation chain or imported uniqueness theorem is used, so the paper is partially circular rather than fully reducible to its inputs. Score 6.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 3 invented entities

The central claims rest on a stack of hand-set parameters (bias, affect weights, loop gains, decays, caution-penalty scale) and domain assumptions (ipsundrum as scaffold, indicator framework as frame, sign convention, forward-model fidelity). No parameter is fit to external data; several are chosen ad hoc to produce the target behaviors. The invented entities (affect proxy, ipsundrum loop, efference copy) are computational abstractions with no independent falsifiable handle — consistent with the paper's own disclaimer, but it means the 'dissociations' are properties of the authors' design choices.

free parameters (8)
  • sensor bias b = 0.5
    Chosen to map signed I_t ∈ [-1,1] into the [0,1] sensor range (Eq. 1); sets the operating point of the recurrence and affects when N_s crosses the confirmation threshold.
  • affect score weights (w_v, w_a, w_s, w_bb) = (2.0, -1.2, -0.8, -0.4)
    Hand-set 'headline affect-coupled' weights in Eq. 8; these directly determine scenic-preference stability and caution behavior.
  • arousal-gated caution penalty (forward moves) = not stated in text
    Ad hoc scoring term 'to link high arousal to avoidance'; the pain-tail tail-duration result (90 vs 5) is produced by this engineered term.
  • hazard-touch penalty scale w_haz = 0.10
    Hand-set; the navigation safety results (0 hazard contacts for the affect variant in CorridorWorld) depend on it.
  • integrator decay d and efference decay d_e = in results/paper/params table.tex (not printed)
    Set the timescale of N_s persistence (Eqs. 3-5); the persistence dissociation rests on these values.
  • loop gain g_eff, scale h, precision π_t = in params table (not printed)
    Set effective recurrence strength α_eff = d + (1−d)(g_eff h π_t) (Eq. 7); the lesion-dependent AUC drops scale with these.
  • optional noise ϵ_t = 0 in headline runs
    All headline results are deterministic given seed; bootstrap intervals reflect seed variation only, not process noise.
  • novelty/curiosity bonus scale = not stated
    Used in the corridor familiarity manipulation and exploratory play; the scan-event and Δscenic-entry results depend on it.
axioms (6)
  • domain assumption Humphrey's ipsundrum narrative (self-monitoring → privatization → re-entrant loop) is treated as a design scaffold, not as a biological claim
    Model Summary and Discussion; the paper is explicit that the operationalization is minimal and not a realization of Humphrey's full theory.
  • domain assumption Butlin et al.'s indicator framework — theory-derived features shift credence; behavioral markers alone are insufficient — is accepted as the methodological frame
    Introduction/Terminology; the entire paper's framing and 'non-claim' structure rest on accepting this framework.
  • domain assumption Signed sensory evidence I_t with positive = noxious and negative = scenic/beneficial is a valid abstraction of the environments
    Model Summary and Eq. 8; the corridor preference dissociation depends on this sign convention.
  • domain assumption The one-step forward model is an accurate model of environment dynamics ('model-aligned forward dynamics'), so planning-time rollouts faithfully predict real consequences
    Policy section; the ablation fidelity and all action-selection results assume the forward model matches the world simulator.
  • ad hoc to paper The sign-asymmetric treatment (non-affect rectifies negative I_t; affect benefits from it) is a legitimate way to isolate affect coupling
    Model Summary; the face-value 'affect vs non-affect' contrast bundles this asymmetry with the affect proxy — acknowledged in Discussion as 'value-shaped'.
  • domain assumption The pain-tail protocol (force one hazard touch, remove hazard, hold state fixed, record planned actions) measures post-stimulus persistence and planned caution rather than ongoing evasion
    Pain-Tail section; the 200-step fixed-state design assumes the tail duration reflects internal state rather than continued environmental threat.
invented entities (3)
  • Affect proxy nodes N_v (valence), N_a (arousal), N_i (body-budget interoceptive sensor) no independent evidence
    purpose: Provide affect readouts that enter the action score (Eq. 8) and modulate recurrence gain/precision
    Defined entirely by the paper's own homeostatic budget model; no independent empirical handle — the authors call it a 'bookkeeping abstraction'.
  • Ipsundrum recurrent loop over sensory salience N_s (feedback E_{t-1} with gain g_eff and precision π_t) no independent evidence
    purpose: Sustain post-stimulus sensory persistence — the key lesion-dependent signature
    An engineered scalar recurrence, not a validated neural or phenomenal mechanism; the paper explicitly disclaims realization of Humphrey's theory.
  • Efference-copy sensor N_e gating loop continuation no independent evidence
    purpose: Stage D gating: the percept script keeps looping while N_e (filtered motor-command magnitude) is above threshold, producing an attractor-like settling regime
    An internally defined monitor signal with no external validation; it directly couples the motor command to loop persistence.

pith-pipeline@v1.3.0-alltime-deepseek · 10335 in / 21069 out tokens · 187510 ms · 2026-08-02T20:25:16.907125+00:00 · methodology

0 comments
read the original abstract

Indicator-based approaches to machine consciousness recommend mechanism-linked evidence triangulated across tasks, supported by architectural inspection and causal intervention. Inspired by Humphrey's ipsundrum hypothesis, we implement ReCoN-Ipsundrum, an inspectable agent that extends a ReCoN state machine with a recurrent persistence loop over sensory salience $N^s$ and an optional affect proxy reporting valence/arousal. Across fixed-parameter ablations (ReCoN, Ipsundrum, Ipsundrum+affect), we operationalize Humphrey's qualiaphilia (preference for sensory experience for its own sake) as a familiarity-controlled scenic-over-dull route choice. We find a novelty dissociation: non-affect variants are novelty-sensitive ($\Delta$scenic-entry = 0.07). Affect coupling is stable ($\Delta$scenic-entry = 0.01) even when scenic is less novel (median {$\Delta$novelty $\approx$ -0.43). In reward-free exploratory play, the affect variant shows structured local investigation (scan events 31.4 vs. 0.9; cycle score 7.6). In a pain-tail probe, only the affect variant sustains prolonged planned caution (tail duration 90 vs. 5). Lesioning feedback+integration selectively reduces post-stimulus persistence in ipsundrum variants (AUC drop 27.62, 27.9%) while leaving ReCoN unchanged. These dissociations link recurrence $\rightarrow$ persistence and affect-coupled control $\rightarrow$ preference stability, scanning, and lingering caution, illustrating how indicator-like signatures can be engineered and why mechanistic and causal evidence should accompany behavioral markers.

Figures

Figures reproduced from arXiv: 2602.23232 by Aishik Sanyal.

Figure 1
Figure 1. Figure 1: Familiarity-controlled corridor preference. Scenic-entry rates under novelty competition. Non-affect variants increase [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Exploratory play. The affect variant shows more scan events and stronger limit-cycle structure without high-entropy [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Pain-tail assay. After hazard contact then removal, ipsundrum variants show non-zero post-stimulus [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Causal lesion. Lesioning feedback+integration reduces post-stimulus persistence ( [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Prosociality by Coupling, Not Mere Observation: Homeostatic Sharing in an Inspectable Recurrent Artificial Life Agent

    cs.MA 2026-04 unverdicted novelty 6.0

    Helping behavior arises in a homeostatic recurrent agent when another's need is routed into self-regulation, but not when the agent only observes the partner.

  2. Prosociality by Coupling, Not Mere Observation: Homeostatic Sharing in an Inspectable Recurrent Artificial Life Agent

    cs.MA 2026-04 unverdicted novelty 6.0

    Homeostatic coupling, not mere partner observation, produces food-sharing in a self-directed recurrent ALife agent with no partner-welfare reward.

Reference graph

Works this paper leans on

3 extracted references · 2 linked inside Pith · cited by 1 Pith paper

  1. [2]

    Brohan, A.; Brown, N.; Carbajal, J.; Chebotar, Y .; Chen, X.; Choromanski, K.; Ding, T.; Driess, D.; Dubey, A.; Finn, C.; et al

    What Is a Cognitive Map? Organizing Knowledge for Flexible Behavior.Neuron, 100(2): 490–509. Brohan, A.; Brown, N.; Carbajal, J.; Chebotar, Y .; Chen, X.; Choromanski, K.; Ding, T.; Driess, D.; Dubey, A.; Finn, C.; et al. 2023. RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control. arXiv:2307.15818. Butlin, P.; Long, R.; Bayne, T.;...

    Pith/arXiv arXiv 2023

  2. [2016]

    Driess, D.; Xia, F.; Sajjadi, M

    Organizing Conceptual Knowledge in Humans with a Gridlike Code.Science, 352(6292): 1464–1468. Driess, D.; Xia, F.; Sajjadi, M. S. M.; Lynch, C.; Chowdhery, A.; Ichter, B.; Wahid, A.; Tompson, J.; Vuong, Q.; Yu, T.; et al. 2023. PaLM-E: An Embodied Multimodal Language Model. arXiv:2303.03378. Friston, K. 2010. The Free-Energy Principle: A Unified Brain The...

    Pith/arXiv arXiv 2023

  3. [2018]

    Barrett, L

    Vector-Based Navigation Using Grid-Like Represen- tations in Artificial Agents.Nature, 557(7705): 429–433. Barrett, L. F. 2017.How Emotions Are Made: The Secret Life of the Brain. Houghton Mifflin Harcourt. Behrens, T. E. J.; Muller, T. H.; Whittington, J. C. R.; Mark, S.; Baram, A. B.; Stachenfeld, K. L.; and Kurth-Nelson, Z

    2017