Goal Inference and Hypothesis Ranking: ASRA Phase 5 — From Interventions to Scientific Objectives

---

Abstract

Phases 1–3 of the Adaptive Scientific Reasoning Architecture (ASRA) established transition logging, object-centric observation, and directed exploration. Phase 4 added action semantics, transition prediction, hypothesis confirm/refute, and counterfactual lookup — the machinery of intervention reasoning.

None of these layers yet answers: What is this system trying to achieve? Which hidden objective best explains the progress we have seen? What experiment should we run next to distinguish competing explanations?

We describe ASRA Phase 5 as the Goal Inference & Hypothesis Engine: a stack that generates win-condition hypotheses from scene structure and semantic operators, detects progress signals from sparse rewards and structural change, classifies object roles, ranks competing goal explanations, and plans discriminating experiments using Phase 4 uncertainty and counterfactuals. The competition agent embeds a compact GoalHypothesisEngine atop Phase 4's CausalSemanticsEngine; the full research engine is specified for asra-arc/src/asra/goals/.

This article presents the theory, architectural decomposition, and design principles. It specifies what Phase 5 adds and why it is the pivot from exploration and causality toward Phase 6 planning and the Phase 8 Decision Biology bridge.

---

1. The architectural gap Phase 5 closes

ASRA's cumulative cognitive stack:

Phase 1   Experience Engine           — transitions, hashes, cell diffs
Phase 2   Observation Engine          — objects, transforms, rule hypotheses
Phase 3   Navigation & Memory         — exploration graph, visitation, subgoals
Phase 4   Semantics & Causal Inference — action meaning, prediction, counterfactuals
Phase 5   Goal Inference & Hypotheses   — win conditions, progress, experiment design
Phase 6+  Planning, robustness

Phase 1 asks: What happened when we acted?
Phase 2 asks: What structural entities changed?
Phase 3 asks: Where should we go next, given memory?
Phase 4 asks: What does this action do, and how sure are we?
Phase 5 asks: What are we trying to accomplish, and how do we test that belief?

Without Phase 5, an agent that understands interventions still acts without purpose: it knows ACTION3 translates an object but not whether translation serves collection, pattern matching, or hazard avoidance. Phase 4 uncertainty drives semantic testing; Phase 5 directs that testing toward goal discrimination.

flowchart LR
  subgraph P4["Phase 4 — Causality"]
    E[Effect signatures]
    U[Uncertainty]
    CF[Counterfactuals]
  end
  subgraph P5["Phase 5 — Goals"]
    GH[Goal hypotheses]
    PD[Progress detector]
    HR[Hypothesis ranker]
    EP[Experiment planner]
  end
  subgraph Future["Phase 6+"]
    PL[Planning]
    DB[Decision Biology]
  end
  E --> GH
  U --> EP
  CF --> EP
  GH --> HR
  PD --> HR
  HR --> EP
  EP --> PL
  GH --> DB

---

2. Theoretical stance: latent objectives without oracle manuals

Interactive environments — and biological systems — rarely publish their win conditions. ASRA Phase 5 treats goals as latent variables inferred from:

1. Progress events — rewards, level changes, WIN terminals, structural alignment.
2. Operator sequences — Phase 4 semantic labels composed over time.
3. Object roles — which entities behave as agents, targets, tokens, or hazards.
4. Template priors — abstractions bootstrapped from Original ARC and domain adapters.

The epistemic object is a goal hypothesis:

h = (template, preferred_operators, object_roles, preconditions, confidence)

Evidence updates support, refute, and progress_score — analogous to Phase 4's causal hypothesis confirm/refute, but at the task level rather than the action-effect level.

This mirrors Decision Biology at the conceptual layer:

Game:     hidden win condition  ← inferred from perturbation–response paths
Cell:     survival objective    ← inferred from perturbation–response paths (Phase 8)

Phase 5 is the first phase where the objective function itself is uncertain and learned — not just the transition dynamics.

Paradigm	Phase 5 stance
Hand-coded win conditions per game	Rejected — goals are inferred
Reward-only RL	Insufficient — reward is sparse; structure carries signal
LLM goal parsing from instructions	Deferred — no instruction channel in ARC-AGI-3
Full neural goal encoders	Deferred — v1 uses templates + ranking
Planning (Phase 6)	Requires Phase 5 objective first

---

3. Goal templates as scientific hypotheses

Phase 5 v1 uses a template library — explicit, falsifiable win-condition families:

Template	Scientific reading	Example operators
move_to_target	Reach a target state in state space	translate
match_pattern	Minimize structural distance to goal configuration	recolor, transform
collect_tokens	Remove or aggregate discrete markers	delete_object
avoid_hazard	Constrained optimization under danger	translate, no_op
unlock_passage	Enable latent pathway	create_object, recolor
transform_to_goal	Sequential mechanism discovery	multi_cell_transform

Each template is a hypothesis class, not a single hypothesis. Instantiation depends on scene-specific object roles and Phase 4 semantics.

Original ARC provides offline supervision: input/output pairs reveal which template family applies — the analog of known biological endpoint assays that constrain pathway hypotheses without revealing full mechanism.

---

4. Progress detection as evidence accumulation

Scientific reasoning requires distinguishing meaningful change from noise. Phase 5's progress detector aggregates:

Signal type	Source	Hypothesis update
reward	Environment	+weight if semantics match leading hypothesis
level_up	levels_completed	Strong support for progress-aligned hypotheses
win	Terminal WIN	Weak confirmation of leading hypothesis
object_progress	Centroid/bbox movement toward target role	Spatial template evidence
pattern_progress	Phase 2 similarity to inferred output	match_pattern evidence

Progress is not assumed monotonic. Refutation occurs when repeated applications of a hypothesized operator fail to produce expected progress — the same loop as Phase 4 hypothesis refute, elevated to goal level.

---

5. Object roles: from pixels to functional entities

Phase 2 extracts objects; Phase 5 assigns roles. Without roles, "move to target" is undefined.

v1 heuristics (no neural net):

• agent — moves across transitions; unique color.
• target — static focal region; often distinct color at border.
• token — multiple identical instances; count decreases on progress.
• hazard — associated with dead-ends or negative progress.
• key / door — state change after specific semantic operator.

CLEVR provides eval signal for relational roles; ARC-AGI-3 provides interactive stress test.

Roles attach to transitions as metadata.goals.object_roles and condition template instantiation.

---

6. Hypothesis ranking as belief maintenance

Given active hypotheses ({h_1, \ldots, h_n}), Phase 5 maintains scores:

score(h) = w_p · progress_score(h) + w_s · support(h) - w_r · refute(h) + ...

The leading hypothesis guides action selection but does not eliminate exploration — wrong goals are common early in episodes.

Status lifecycle:

active → leading → confirmed (on WIN) / refuted (on counter-evidence) / weak (low support)

Ranking output feeds:

• Kaggle agent reasoning strings (goal=move_to_target).
• Phase 6 planner objective selection.
• Phase 8 pathway hypothesis priors.

---

7. Experiment planning: discrimination over confirmation

Phase 4 asks: Which action reduces semantic uncertainty?
Phase 5 asks: Which action separates hypothesis A from hypothesis B?

Given top hypotheses (h_1, h_2) and candidate action (a):

discrimination(a) = |match(sem(a), h_1) - match(sem(a), h_2)| × uncertainty(s, a)

High discrimination + high uncertainty → priority experiment. This is the game analog of designed biological perturbations that distinguish competing signaling models.

Phase 4 counterfactuals supply predicted alternate outcomes for planning; Phase 5 chooses which counterfactual question matters for goal belief.

---

8. Closing the loop with Phases 1–4

Layer	Phase 5 consumption
Phase 1 transitions	Progress event stream; WIN mining
Phase 2 scenes	Object roles; pattern progress
Phase 2 Original ARC	Template bootstrap
Phase 3 subgoals	level_progress as progress signal
Phase 3 exploration	Fallback when hypotheses weak
Phase 4 semantics	Operator vocabulary for templates
Phase 4 uncertainty	Experiment planner information gain
Phase 4 counterfactuals	Predicted outcomes for discrimination

Kaggle agent scoring (embedded):

score(action) = Phase1–4_terms
              + GOAL_HINT_WEIGHT · action_goal_score(leading, sem)
              + EXPERIMENT_HINT_WEIGHT · discrimination(top-2, sem, u)

Reasoning strings:

ASRA Phase5: ACTION3 | objects=5 | visits=2 | sem=translate conf=0.81 u=0.12 | goal=move_to_target

---

9. Empirical landscape

Phase 5 metrics measure goal reasoning quality, not competition win rate (Phase 6).

Benchmark	Metric
ARC-AGI-3	Leading hypothesis stability; progress correlation; WIN hindsight rank
Original ARC	Template family classification from input grid
PHYRE	Success-template match; experiments-to-success
CLEVR/CLEVRER	Object role accuracy; CF refute rate

Phase 5 deliberately does not claim Milestone #2 — it builds the objective layer planners require.

---

10. Position in the ASRA research program

Question	Phase 4	Phase 5
Why try action a?	Semantics, uncertainty	+ Goal alignment, discrimination
Unit of task memory	Effect signatures	Ranked goal hypotheses
What is success?	Observed reward	Inferred win condition
Bridge to biology	Intervention form	Latent objective form

From the Decision Biology roadmap:

environment state  →  action  →  next state        (Phase 4)
hidden goal        →  inferred from responses     (Phase 5)
cell objective     →  survival / adaptation         (Phase 8)

Phase 5 introduces objective uncertainty — the missing variable between causal dynamics and adaptive planning.

---

11. Kaggle submission and agent evolution

Version	Tag	Layer added
Phase 4	asra-v0.6-phase4	Causal semantics
Phase 5	asra-v0.7-phase5	Goal hypotheses, progress, experiment hints

The companion notebook writes my_agent.py, self-tests perception + exploration + causality + goals, and emits validation parquet. Full library (goals/ package, ARC batch mining, PHYRE/CLEVR eval) remains specified for offline research; the competition agent carries a minimal sufficient goal stack.

---

12. Open problems and next theory steps

1. Planning (Phase 6) — use leading hypothesis as planner objective; BFS/A* over semantic operators.
2. Conditional goals — same template, different parameters by game context.
3. Neural goal encoders (v2) — when template library saturates.
4. Cross-game goal transfer — meta-templates from Original ARC to unseen ARC-AGI-3 games.
5. Decision Biology (Phase 8) — swap templates for pathway survival objectives; reuse ranking loop on LINCS/scPerturb.

---

13. Conclusion

ASRA Phase 5 is the project's shift from understanding interventions to inferring objectives: goal templates make implicit win conditions explicit; progress detectors turn sparse feedback into evidence; hypothesis ranking maintains belief over competing explanations; experiment planning directs action toward discrimination, not only discovery.

The Phase 5 Kaggle extension is not a new agent philosophy — it is Phase 2 + Phase 3 + Phase 4 + purpose. Semantics still describe what actions do; goals describe why they matter.

Transition-centric adaptive reasoning remains the core; goal inference is how those transitions become meaningful toward an objective — the prerequisite for planning, robustness, and the Decision Biology bridge.

---

Reference notebook (GitHub)

Interactive companion with Phase 2 object-scene hints, Phase 3 exploration memory, Phase 4 causal semantics, and Phase 5 goal hypothesis ranking:

• ASRA Phase 5 — ARC Prize 2026 (ASRA repository)
• SciLayer archive copy

---

References

1. Chollet, F. On the Measure of Intelligence. arXiv (2019).
2. Pearl, J. Causality: Models, Reasoning, and Inference. Cambridge University Press (conceptual lineage).
3. Ilakkuvaselvi Manoharan. Transition-Centric Adaptive Reasoning: ASRA Phase 1 for Interactive Environments. https://sci-layer.vercel.app/articles/transition-centric-adaptive-reasoning-asra-phase-1
4. Ilakkuvaselvi Manoharan. Object-Centric Adaptive Reasoning: ASRA Phase 2. https://sci-layer.vercel.app/articles/object-centric-adaptive-reasoning-asra-phase-2
5. Ilakkuvaselvi Manoharan. Directed Exploration and Episodic Memory: ASRA Phase 3. https://sci-layer.vercel.app/articles/directed-exploration-episodic-memory-asra-phase-3
6. Ilakkuvaselvi Manoharan. Causal Action Semantics: ASRA Phase 4. https://sci-layer.vercel.app/articles/causal-action-semantics-asra-phase-4
7. Ilakkuvaselvi Manoharan. ASRA: Adaptive Scientific Reasoning Architecture. https://github.com/ilakkmanoharan/asra
8. Phase 5 goals implementation — https://github.com/ilakkmanoharan/asra/tree/main/asra-arc/src/asra/goals

---

Related: ASRA Phase 4 · Action semantics inference · Decision Biology · Nature Foundation Models

Correspondence: ilakkmanoharan@gmail.com