# ============================================================================= # needs.py # ----------------------------------------------------------------------------- # Responsible for: The "Individual Value System" of EduMirror -- the # Psychological Need System with 5 major categories and 13 # sub-dimensions, the trait -> expected-value mapping, and the # unmet-need gap computation. # Role in project: This is the motivational core that the Value-driven Planner # (agents.py) reads from and that the Social Value System # (svo.py) aggregates over. It is the direct subject of # Claim 1 of the reproduction. # Assumptions: Follows arXiv:2606.07948 Section 3.3 ("Individual Value System") # and Appendix E.1 / Table 9. All values live on a 0-10 Likert # scale. No LLM calls happen in this file -- it is pure state, so # it can be unit-tested deterministically. # ============================================================================= """Psychological Need System for EduMirror value-driven agents. Paper grounding (arXiv:2606.07948): - Sec 3.3: "We formalize the Value System as a Psychological Need System comprising five major categories, namely Safety, Mental Health, Self-Esteem, Social Belonging, and Meaning and Growth, with 13 sub-dimensions in total. Each need is represented on a Likert scale ranging from 0 to 10." - Sec 3.3: unmet-need gap Delta_t(d) = clip(v*(d) - v_t(d), 0, S_max) - App E.1 + Table 9: the 13 named sub-dimensions, their associated personality traits, and the degree-adverb -> expected-value mapping (slightly->7.5, moderately->8, quite->8.5, extremely->9). Note on naming: the main text (Sec 3.3) and the appendix (E.1) use slightly different labels for two categories -- main text says "Mental Health" and "Self-Esteem", the appendix says "Psychological Health Needs" and "Esteem". We keep the main-text category names as canonical and record the appendix aliases, since the sub-dimension membership is identical in both. """ from __future__ import annotations import random from dataclasses import dataclass, field # ----------------------------------------------------------------------------- # The taxonomy. This literal IS the claim under test: 5 categories, 13 # sub-dimensions, each sub-dimension bound to the personality trait that # Table 9 associates with it. # ----------------------------------------------------------------------------- #: Likert scale bounds for every need dimension (paper Sec 3.3: "0 to 10"). VALUE_MIN, VALUE_MAX = 0.0, 10.0 #: S_max, the clip ceiling on a single dimension's unmet-need gap. The paper #: writes clip(v* - v, 0, S_max) without pinning S_max numerically; since both #: v* and v live on [0, 10], a gap can never exceed 10, so S_max = 10 is the #: tightest bound that leaves the paper's formula unchanged. S_MAX = 10.0 #: Appendix E.1 degree-adverb -> expected-value (v*) mapping. The paper #: initializes v* high, in [7.5, 9.0], to encode "desired state of well-being": #: a higher expectation makes the agent more sensitive to a deficit. DEGREE_TO_EXPECTED = { "slightly": 7.5, "moderately": 8.0, "quite": 8.5, "extremely": 9.0, } #: The Psychological Need System: category -> ordered sub-dimensions. #: Sub-dimension names and their trait bindings come from Table 9; the #: category groupings come from Appendix E.1's enumerated list. NEED_TAXONOMY: dict[str, list[str]] = { "Safety": [ "psychological safety", "emotional safety", ], "Social Belonging": [ "group acceptance", "support system", "sense of superiority", ], "Self-Esteem": [ "self worth", "sense of respect", ], "Meaning and Growth": [ "sense of meaning", "sense of control", "passion and motivation", ], "Mental Health": [ "emotional stability", "emotional wellbeing", "psychological resilience", ], } #: Appendix E.1 uses these alternative labels for two of the categories. #: Recorded so the reproduction can be checked against either wording. CATEGORY_ALIASES = { "Mental Health": "Psychological Health Needs", "Self-Esteem": "Esteem", } #: Table 9: mapping between each psychological need and its associated trait. NEED_TO_TRAIT: dict[str, str] = { "psychological safety": "Timid", "emotional safety": "Emotionally Sensitive", "group acceptance": "Sociable", "support system": "Dependent", "sense of superiority": "Competitive", "self worth": "Reputation-conscious", "sense of respect": "Ego-driven", "sense of meaning": "Spiritual", "sense of control": "Possessive", "passion and motivation": "Passionate", "emotional stability": "Emotionally Stable", "emotional wellbeing": "Hedonistic", "psychological resilience": "Resilient", } #: Flat ordered list of all 13 sub-dimensions. ALL_NEEDS: list[str] = [d for dims in NEED_TAXONOMY.values() for d in dims] #: Reverse index: sub-dimension -> its parent category. NEED_TO_CATEGORY: dict[str, str] = { d: cat for cat, dims in NEED_TAXONOMY.items() for d in dims } def qualitative_description(value: float) -> str: """ Convert a raw 0-10 need score into a natural-language band label. Args: value: The current need score on the 0-10 Likert scale. Returns: A short adjective phrase describing the satisfaction level. Why: Appendix E.1 is explicit that "large language models struggle to interpret raw numerical values", so EduMirror runs a "qualitative description" step that textualizes v_t before handing it to the planner. Reproducing that step matters: feeding bare numbers into the prompt would be a different (and per the paper, worse) architecture. The band edges are our choice -- the paper describes the mechanism but does not publish its thresholds. """ if value < 2.0: return "severely unmet" if value < 4.0: return "largely unmet" if value < 6.0: return "partially met" if value < 8.0: return "mostly met" return "fully satisfied" @dataclass class NeedState: """ The evolving psychological need state of a single EduMirror agent. Attributes: current: Sub-dimension -> current value v_t(d), on [0, 10]. expected: Sub-dimension -> expected value v*(d), on [7.5, 9.0]. traits: The sampled " " phrases describing this agent, e.g. "extremely Timid". Used to build the persona prompt. Why: The paper separates *current* from *expected* per dimension so that the unmet-need gap (the planner's objective) is personality-relative rather than absolute: two agents at the same v_t feel different degrees of deprivation if their v* differ. Collapsing these into one number would destroy the trait-sensitivity the paper's Figure 5 depends on. """ current: dict[str, float] expected: dict[str, float] traits: dict[str, str] = field(default_factory=dict) def gap(self, dim: str) -> float: """ Unmet-need gap for one dimension: Delta_t(d) = clip(v*(d) - v_t(d), 0, S_max). Args: dim: Sub-dimension name (must be one of ALL_NEEDS). Returns: Non-negative gap in [0, S_MAX]. Zero when the need is at or above its expectation. Why: This is Eq. (unnumbered) in Sec 3.3 verbatim. Clipping at 0 encodes "a need that is over-satisfied creates no drive" -- an agent whose safety exceeds its expectation is not motivated to seek more safety. """ raw = self.expected[dim] - self.current[dim] return min(max(raw, 0.0), S_MAX) def gaps(self) -> dict[str, float]: """ Unmet-need gaps across all 13 dimensions. Args: None. Returns: Sub-dimension -> gap, for every dimension in ALL_NEEDS. Why: The planner and the Social Value System both consume the full gap vector (Delta_t), so materializing it once keeps them consistent. """ return {d: self.gap(d) for d in ALL_NEEDS} def category_means(self) -> dict[str, float]: """ Mean current value per major category. Args: None. Returns: Category name -> mean of its sub-dimensions' current values. Why: Figures 5 and 6 of the paper plot psychological dynamics at the 5-category level, not the 13-dimension level, so the reproduction needs this aggregation to produce comparable plots. """ return { cat: sum(self.current[d] for d in dims) / len(dims) for cat, dims in NEED_TAXONOMY.items() } def describe(self) -> str: """ Render the full need state as the textual block shown to the planner. Args: None. Returns: A multi-line string, one line per category, listing each sub-dimension's qualitative band and its unmet gap. Why: Implements the Appendix E.1 "qualitative description" process that precedes every planning step. Grouping by category keeps the prompt compact and mirrors how the paper describes the taxonomy. """ lines = [] for cat, dims in NEED_TAXONOMY.items(): parts = [] for d in dims: band = qualitative_description(self.current[d]) gap = self.gap(d) # Surfacing the gap (not just the band) is what lets the planner # rank candidate actions by need-gap reduction, per Sec 3.3. parts.append(f"{d}: {band} (unmet gap {gap:.1f})") lines.append(f"- {cat}: " + "; ".join(parts)) return "\n".join(lines) def apply_deltas(self, deltas: dict[str, float]) -> None: """ Update current need values in place, clamped to the Likert range. Args: deltas: Sub-dimension -> signed change to apply. Unknown keys are ignored so a noisy LLM update cannot inject new dimensions. Returns: None. Mutates `self.current`. Why: Appendix E.1's update step integrates (a_t, o_t, v_{t-1}, d_{t-1}) into v_t. We let the LLM propose per-dimension deltas rather than absolute values, because deltas keep the update anchored to the previous state and make a malformed response degrade gracefully (a missing dimension simply does not move) instead of resetting the agent's history. """ for dim, delta in deltas.items(): if dim not in self.current: continue # ignore hallucinated dimensions self.current[dim] = min(max(self.current[dim] + delta, VALUE_MIN), VALUE_MAX) def init_need_state(rng: random.Random) -> NeedState: """ Initialize an agent's need state by sampling traits and values, per Appendix E.1. Args: rng: Seeded RNG. Passed in (not module-global) so a whole simulation is reproducible from a single seed. Returns: A NeedState with expected values in [7.5, 9.0] derived from sampled degree adverbs, and current values sampled uniformly from [0, 10]. Why: The paper: "At initialization, adjectives and degree adverbs are randomly selected to establish these personal expected values, while the initial current scores v_0 are randomly sampled within [0, 10]." The wide v_0 range is deliberate -- it is what produces the "higher initial values enhance resilience, lower values increase volatility" contrast the paper reports in Figure 5. """ expected: dict[str, float] = {} current: dict[str, float] = {} traits: dict[str, str] = {} for dim in ALL_NEEDS: degree = rng.choice(list(DEGREE_TO_EXPECTED)) expected[dim] = DEGREE_TO_EXPECTED[degree] current[dim] = rng.uniform(VALUE_MIN, VALUE_MAX) traits[dim] = f"{degree} {NEED_TO_TRAIT[dim]}" return NeedState(current=current, expected=expected, traits=traits) def init_need_state_from_profile( rng: random.Random, degree_by_need: dict[str, str] | None = None, baseline_current: float | None = None, ) -> NeedState: """ Initialize a need state with explicit control over traits and starting values. Args: rng: Seeded RNG, used for any dimension not pinned by `degree_by_need`. degree_by_need: Optional sub-dimension -> degree adverb (one of DEGREE_TO_EXPECTED). Dimensions left out are sampled randomly. baseline_current: If given, every current value starts here instead of being sampled from [0, 10]. Returns: A configured NeedState. Why: Two experiments need determinism that `init_need_state` cannot give. Case Study 1 contrasts "high initial state" vs "low initial state" victims (Figure 5), which requires pinning v_0. The intervention experiments (Figure 6) require "20 bullying scenarios with identical initial settings", which requires pinning both v_0 and v*. This is the seam that makes those controlled comparisons possible. """ state = init_need_state(rng) if degree_by_need: for dim, degree in degree_by_need.items(): if dim in state.expected: state.expected[dim] = DEGREE_TO_EXPECTED[degree] state.traits[dim] = f"{degree} {NEED_TO_TRAIT[dim]}" if baseline_current is not None: for dim in state.current: state.current[dim] = baseline_current return state