Hippocampal Valence Asymmetry: BLA vs VTA Pathways

Registered: 2026-04-14 Claims: MECH-232, MECH-233, ARC-057 Literature: targeted_review_hippocampal_valence_asymmetry (5 entries)

The Problem

REE’s residue field encodes harm/threat valence directly in hippocampal terrain – the agent avoids locations with high residue. This maps cleanly to the BLA-hippocampal pathway (Jimenez 2018: ventral CA1 “anxiety cells”; Redondo 2014: BLA engrams resist valence reversal). The mechanism is explicit tagging: certain hippocampal cells encode “this context is dangerous” and drive avoidance via vCA1->LHA projections.

The architectural question: does approach/wanting valence enter hippocampal terrain the same way? Is there a symmetric “wanting gradient” that mirrors the harm residue field? The literature says no.

The Asymmetry

Threat pathway (BLA -> hippocampus)

Dedicated valence-coding cells in ventral CA1 (Jimenez 2018)
Explicit avoidance tags, fast, one-shot (contextual fear conditioning)
BLA engrams cannot be valence-switched (Redondo 2014)
Possibly compressed representation (fast detection, not fine spatial detail)
Maps directly to REE residue field

Approach pathway (VTA -> hippocampus)

No equivalent “reward cells” or “wanting cells” found in hippocampus (Jimenez 2018)
Dopaminergic modulation enhances LTP at reward-associated locations (Lisman & Grace 2005)
Place cells reorient to represent reward-relevant spatial features; D1 blockade prevents reorientation and causes perseveration (Retailleau & Morris 2018)
Reward-associated items have stronger, more durable memory traces (Wittmann 2005)
The mechanism is representational expansion, not valence tagging

MECH-232: DA-Mediated Representational Expansion

Dopaminergic modulation at reward locations does not write a positive valence tag. It enhances the fidelity, stability, and associative richness of place representations. The result: reward-associated locations have more detailed, more durable, more easily reactivated maps. The approach gradient is implicit in representational quality.

This means the hippocampal terrain at reward locations is effectively “expanded” – it contains more distinguishable states, more fine-grained spatial features, more navigable structure in representational space.

MECH-233: Mechanistic Asymmetry

The two pathways produce behaviorally similar outputs (avoidance of threat, approach toward reward) through mechanistically distinct encoding:

Feature	Threat (BLA)	Approach (VTA-DA)
Encoding type	Explicit valence tag	Representational expansion
Cell type	Dedicated anxiety cells	Enhanced place cells
Speed	Fast, one-shot	Gradual, consolidation-dependent
Reversibility	BLA engrams resist switching	Expansion decays when DA ceases
Map effect	Tags location	Expands representation
Curiosity interaction	Wall (avoid)	Open field (explore)

ARC-057: Curiosity-Approach Emergence

The architectural commitment: no explicit wanting gradient is required in hippocampal terrain. Approach behavior emerges from two interacting mechanisms:

DA-mediated representational expansion (MECH-232): reward locations have more navigable representational structure
Information-seeking drive (curiosity): the agent explores available structure

The curiosity drive does the same thing everywhere – it follows available representational structure. But “available” is asymmetric because dopamine shaped the map. Reward-associated regions have more structure to explore, so the agent spends more time there. An observer infers an attractive gradient, but there is no gradient – there is more map.

Sub-predictions

Wanting-as-maintenance: DA re-sharpening on each reward encounter refreshes the representational expansion. The map stays expanded as long as the reward contingency holds. When rewards stop and DA modulation ceases, expansion decays and approach behavior extinguishes.
Wanting-as-sensitivity: Already-expanded representations encode incoming resource/goal gradients at higher fidelity. The agent detects goal signals at greater distance because its map of reward-associated areas has higher resolution.
Double dissociation: Ablating DA modulation degrades approach but not avoidance. Ablating explicit valence tags degrades avoidance but not approach. Both components are independently necessary for their respective behaviors.

Substrate Constraint: Environmental Complexity

ARC-057 has a minimum environmental complexity requirement that the current CausalGridWorld cannot satisfy. The curiosity-approach mechanism depends on representational expansion capturing genuinely additional information at reward locations. In the real world, this works: any location has near-fractal complexity where zooming in reveals more structure (textures, objects, micro-features, social history). A richer representation at a reward location captures real information that a coarser one misses.

In a grid world, a cell is a cell. There is nothing more to discover at higher resolution. Expanding the representation at grid cell (3,4) just encodes the same sparse features more redundantly. The curiosity drive has nothing additional to explore, so the approach-emergence mechanism cannot operate.

This does NOT affect the threat/avoidance side. Explicit valence tagging (the BLA pathway, the harm residue field) works regardless of environmental richness – tagging a location as dangerous requires no additional spatial detail. The asymmetry between the two pathways means they have different substrate requirements for testing.

Faithful testing of ARC-057 requires either:

A richer environment with location-dependent feature complexity
A conceptual/mind-map space where “locations” have variable information density
Or acceptance that ARC-057 is a theoretical claim grounded in neuroscience literature, testable only when the substrate supports sufficient environmental richness

Implications for REE Architecture

The harm residue field is architecturally correct as-is – it implements the BLA explicit-tagging pathway.
A symmetric wanting gradient field should NOT be added – it would be biologically inaccurate and architecturally redundant.
Instead, the hippocampal module needs DA-modulated representational resolution: a mechanism by which reward encounters enhance the spatial detail encoded at those locations.
The curiosity/exploration drive (already present or planned) then produces approach behavior as an emergent property.
This resolves the “wanting-in-terrain” question for ARC-007: the approach signal IS in the terrain, but as enhanced representational quality, not as a gradient.
Proxy mechanism policy: until a richer environment is available, approach behavior in V3/V4 grid world experiments uses proxy mechanisms (explicit gradient, DA-modulated priority, etc.). These experiments must be tagged approach_proxy and carry evidence quality notes flagging the proxy. Results may not generalize to ARC-057-enabled substrate. See docs/roadmap.md ARC-057 gate section.