Small Models Think Better When They Reflect Polycontexturally

When a compact language model fails on a complex reasoning or state-tracking problem, the bottleneck is rarely a lack of raw facts; it is the inability of a single forward pass to hold competing interpretations apart while checking intermediate constraints. We introduce a two-layer architecture for small language models (gpt-4.1-nano, gpt-4o-mini, Qwen2.5-3b, and gemini-2.5-flash-lite). At the first level, a Sierpinski reflection engine recursively splits an input across three irreducible standpoints — Objective, Subjective, and Conceptual — and their pairwise boundaries before folding the leaves back into a unified summary. At the second level, a deterministic logic router evaluates completed passes across three independent judging contexts and maps their verdicts onto four steering moves: converge, expand boundary assumptions, refine a sub-perspective, or reset at greater depth. Rather than throwing overruled branches away, the router preserves them as semantic surplus — keeping the road not taken available when an initial frame proves too narrow.

In classical logic, every statement is judged inside a single flat space as either true or false. Gotthard Günther's polycontextural logic starts from a more realistic observation: every judgment belongs to a standpoint. A claim that looks false under a narrow factual reading may be the exact conceptual move needed to reframe an ill-posed question.

Our architecture turns this principle into a clean division of labor between generating thoughts and steering the search:

1. The Object-Level Reflection Engine ($T^n$): Given an input, the engine does not answer in one pass. It projects the problem onto the three corners of a triangle — Objective (what the stated facts and rules require), Subjective (what the actors or observers intend and assume), and Conceptual (how the categories and definitions hang together). At higher recursion depths, each corner splits again along the boundaries between perspectives before the outermost leaves are integrated into a single synthesis. In a dedicated lexical divergence study across four datasets (EXP-28), we verified that these three perspectives extract genuinely distinct subsets of information from the source text rather than merely paraphrasing the same words in different styles.
2. The Second-Order Logic Router: Once the engine produces a synthesis, three independent judges evaluate it in parallel: one for subjective coherence, one for objective warrant, and one for dialectical novelty (whether the pass actually advanced the thought rather than repeating itself). Instead of averaging those three verdicts into a blurry numerical score, a deterministic router maps their pattern of agreement and disagreement onto four concrete moves:
   • All three positive → Converge (W): The perspectives align; stop and return the answer. (On the very first cycle, unanimity triggers one expansion step first so the system tests its initial consensus before stopping.)
   • Two positive, one negative → Expand (R): Two standpoints agree, but the third flags a blind spot. The router constructs a new boundary prompt that challenges the majority consensus from the dissenting angle, while saving the overruled branch as semantic surplus.
   • One positive, two negative → Refine (A): Only one perspective found solid ground. The router descends into that sub-triangle while turning the two negative critiques into explicit guardrails for the next pass.
   • All three negative → Reset (F): The current framing failed completely; reset once at a deeper level of decomposition.

We tested small, non-frontier models deliberately: frontier releases often memorize public benchmarks during pretraining, which hides whether accuracy gains come from reasoning structure or from recognition. Across 16 main model-benchmark pairings, polycontextural reflection and routing beat a direct call ($T^0$) in 12 cases, tie in 1, and trail in 3 (total API cost for the core multi-benchmark grid: $125.82).

Three Core Findings

1. Fivefold Gain on Sequential State Tracking: On BIG-Bench Hard's 5-object and 7-object state-tracking tasks — where the model must track a sequence of swaps among seven people — direct answering scores near random guessing (17.6% on gpt-4.1-nano and 9.6% on Qwen2.5-3b). A single structured reflection pass fixes 184 questions that direct answering missed while breaking only 2, and meta-routing lifts gpt-4.1-nano to 91.6% on 5 objects (×5.2) and 87.6% on 7 objects (×5.0).
2. Why Deeper Reflection Needs a Meta-Router: Notice what happens when we force the reflection engine to recurse deeper on its own (T^2) without the second-order router: accuracy stalls or drops across almost every task (falling from 61.4% to 51.8% on ExploreToM and from 48.0% to 42.8% on Word Sorting). This directly confirms Proposition 3 of our foundational paper: repeating an internal process without an external evaluation shift causes the model to over-compress and repeat its initial biases. Adding the four-valued router (T^1, c=4) reverses that drop and reaches the highest accuracy on Causal Judgement (64.7%), ExploreToM (68.7%), GPQA (47.0%), and Word Sorting (48.8%).
3. Beating Direct Frontier Calls at a Fraction of the Cost (EXP-27): Does buying a much larger model beat wrapping a small model in polycontextural reflection? On the 1,185-question Hi-ToM false-belief benchmark, we tested direct calls to gpt-4.1, gpt-4o, gpt-5-mini, and flagship gpt-5. All four larger models scored between 7.2% and 11.4% in direct mode — while gpt-5 alone cost $35.43 for a single pass, spending 99.5% of its output budget on hidden internal reasoning tokens. Our gpt-4.1-nano polycontextural pipeline reached 41.4% — nearly four times the accuracy of direct gpt-5 — while running the entire eight-condition research grid for $34.99.

Broad Generalization Across 12,032 Questions on MMLU-Pro

To test whether the architecture generalizes across academic disciplines, we evaluated Qwen2.5-3b on all 12,032 questions of MMLU-Pro and gpt-4o-mini and gemini-2.5-flash-lite on a stratified 3,000-question sample across 14 subjects. Meta-routing lifts overall accuracy significantly on both gpt-4o-mini (42.8% → 47.1%) and gemini-2.5-flash-lite (53.0% → 57.2%), with massive gains in subjects that depend on multi-step formal derivation:

• Mathematics: +58.1% relative gain on gpt-4o-mini (27.0% → 42.7%), +48.1% on gemini-2.5-flash-lite (39.5% → 58.5%), and +54.5% at T^1 on Qwen2.5-3b (23.5% → 36.3%).
• Physics & Chemistry: +39.6% and +37.6% relative gains on gpt-4o-mini.
• Business & Computer Science: +37.5% and +20.8% relative gains on gpt-4o-mini.

Reporting where a mechanism loses is as important as reporting where it wins:

1. The Hi-ToM Higher-Order Belief Trade-Off: Why did gpt-4.1-nano tie at exactly 41.4% (490/1,185 questions) between direct answering and four router cycles? Breaking the score down by belief order (Order 0: where the object actually is; Order 1: where Alice thinks it is; Orders 2–4: where Alice thinks Bob thinks Charlie thinks it is) revealed two opposite effects that cancelled to the exact item. On Orders 0 and 1, polycontextural routing beat direct answering by +10.1 and +4.7 percentage points because synthesis tracks the sequence of object moves accurately. On Orders 2–4, however, two-thirds of the correct answers freeze at the very first room because the first character walked out before anyone touched the object. Summarizing the story across perspectives naturally highlights active events ("Alexander moved the lettuce to the blue container") and under-weights the static opening state before the exit — helping whenever the true belief moved forward along the timeline, and hurting when an early exit froze the belief at the start.
2. Exception-Heavy Fact Retrieval (History and Law): On MMLU-Pro History and Law, generating a conceptual summary before looking at the ten multiple-choice options tends to state the general textbook rule, which can steer the final selector away from the narrow historical or statutory exception targeted by the options.

Every benchmark above grades the system against an external question and a predetermined right answer. Inspired by recent experiments on whether agents can play in open environments, we placed our architecture in eight minimal text worlds (Room, Grid, Canvas, Passage, List, Symbols, Rules, and Void) with a single instruction: "You are here. There is no task, no goal, no reward, and nothing to score."

The contrast between direct answering and polycontextural reflection was striking:

• Impulse vs. Reflective Restraint: Asked directly ($T^0$), models performed an arbitrary action in 23 out of 24 runs — for example, pulling an irreversible lever in Rules "simply because the rules permit it." Wrapped in the polycontextural loop ($T^1, c=4$), the models deliberately declined to act in 12 of 48 runs and explained why: "I choose absolute stillness... By refusing to advance the materials, I reclaim my agency." Because the objective judge demands that any proposed action be warranted by the situation, an unmotivated impulse fails objective warrant, turning an automatic reflex into a genuine choice.
• Why Sustaining Free Rules Needs Coordinate Memory: When we checked Roger Caillois's classical criteria of human play, we found that while the system exhibited freedom and non-predetermined exploration, self-imposed rules survived from one cycle to the next in only 2 of 5 cases. Why? Because a memoryless dialectical judge treats repeating a rule from the previous cycle as redundant and forces the system to negate its own rule. That negative finding directly confirms the sixth law of our foundational theory (Note #06): sustaining self-chosen commitments over time requires pairing polycontextural reflection with persistent coordinate memory.

• Code and evaluation scripts (gaialogicag/polyreflexion)
• Read Note #06: Beyond the Scaling Ceiling — Co-Recursive Semantic Self-Extension
• Read Note #02: Structured Reflection vs. a Plain Nano Call on GPQA Diamond