1. The Problem

When a researcher asks "How are these two papers connected?", the standard approach is a single LLM prompt. This fails structurally:

1. Premature convergence: LLMs optimize for one plausible narrative, not exhaustive coverage
2. Path of least resistance: Methodological and citation connections (easy) drown out paradigm and synthesis connections (valuable)
3. No stopping criterion: The model halts when it "feels done," not when coverage is complete
4. Context overflow: Full arXiv PDFs (20-50 pages) exceed context windows; naive chunking loses cross-section connections

This is not a model capability problem. It's a process discipline problem.

2. The Insight: TOC as Operating Logic

Goldratt's Theory of Constraints states: every system has exactly one binding constraint, and improving non-constraints yields negligible gains.

Applied to connection-finding:

The Drum-Buffer-Rope mechanism schedules token flow:

• Drum: The active constraint sets the pace
• Buffer: Partial extractions protect the Drum from starvation
• Rope: Token signal releases upstream work at Drum's consumption rate

3. Paper Ingestion via RLM

3.1 The Context Problem

Full arXiv PDFs present a context challenge:

• Typical paper: 20-50 pages
• At ~4k tokens/page: 80k-200k tokens per paper
• Two papers: 160k-400k tokens just for input
• Most LLMs can't handle this efficiently

3.2 RLM Solution

Recursive Language Models (Zhang et al., 2026) enable the LM to programmatically examine, decompose, and recursively call itself over its input. Instead of:

# Traditional: context overflow
llm.completion(prompt + full_paper_text, model)

We use:

# RLM: programmatic decomposition
rlm.completion(prompt, model)  # LM can navigate papers as variables

The RLM paradigm treats paper content as a variable in a REPL environment. The LM can:

1. Examine: Query specific sections/pages on demand
2. Decompose: Break papers into dimension-relevant chunks
3. Recursively call: Launch sub-LM calls for deep analysis

4. The 15 Connection Dimensions

We formalize 15 distinct dimensions, organized by TOC constraint types:

Physical (Tangible Shared Artifacts)

Policy (Methodological Agreements)

Paradigm (Conceptual Relationships)

D15 is the highest-value dimension and the typical Drum.

5. Architecture

5.1 State

@dataclass
class State:
    papers: tuple[Paper, Paper]       # RLM-accessible paper objects
    connections: list[Connection]     # discovered
    coverage: dict[str, float]        # dimension -> [0,1]
    active_constraint: str            # current bottleneck
    buffer: list[PartialResult]       # DBR buffer
    iteration: int

5.2 The Five-Step Loop

def toclink(paper_a: Paper, paper_b: Paper) -> list[Connection]:
    S = State(papers=(paper_a, paper_b))
    
    while min(S.coverage.values()) < THRESHOLD:
        # 1. IDENTIFY
        S.active_constraint = min(S.coverage, key=S.coverage.get)
        
        # 2. EXPLOIT (via RLM for full-text access)
        new = exploit(S.active_constraint, S.papers)
        S.connections.extend(new)
        S.coverage[S.active_constraint] = update_coverage(new)
        
        # 3. SUBORDINATE
        for d in DIMENSIONS - {S.active_constraint}:
            S.buffer.append(partial_extract(d, S.papers))
        
        # 4. ELEVATE (if stuck)
        if coverage_stalled(S):
            elevate(S.active_constraint, S)
        
        # 5. REPEAT (implicit)
    
    return deduplicate(S.connections)

6. Implementation

No LangChain. No LlamaIndex. No vector DB. RLM handles context.

7. Example Run

Paper A: Attention Is All You Need (Vaswani 2017)
Paper B: Flash-KMeans (arXiv 2603.09229)

D15 synthesis was generated on iteration 3 after RLM elevation deep-dived into both papers' methodology sections. A single-pass approach never produced it.

8. Why This Works

8.1 The Throughput Discipline

Naive prompting is like a factory where every machine runs at uncoordinated capacity—the bottleneck gets no special attention and leaves work incomplete.

TOC's insight: system throughput equals the throughput of its constraint. The worst-covered dimension bounds overall quality. TOCLINK forces this dimension to receive disproportionate attention every cycle.

8.2 Breaking the Policy Constraint

The LLM's prior is a policy constraint in Goldratt's sense: it strongly favors D6-D7 (methodological) and underproduces D11-D15 (paradigm). This is invisible to the model—it takes its own behavior for granted.

TOCLINK breaks this by:

1. Explicit coverage scoring exposes the constraint
2. Forced elevation overrides the default generation policy
3. RLM deep-dive enables exhaustive section-by-section analysis
4. DBR scheduling prevents early termination

9. Conclusion

TOCLINK demonstrates that importing an industrial operations framework into AI agent design yields measurable benefits. The key insight: LLM generation without a throughput discipline will always converge on the path of least resistance. TOC's Five Focusing Steps provide exactly the corrective: identify the constraint, exploit it, subordinate everything else, repeat.

RLM integration ensures full-text coverage without context overflow—the LM can programmatically navigate papers as variables, launching sub-calls for deep analysis only when needed.

The result: a ~180-line agent that discovers synthesis hypotheses—novel research directions combining two papers—that single-pass prompting never surfaces.

References

• Goldratt, E. (1984). The Goal. North River Press.
• Zhang, A.L., Kraska, T., Khattab, O. (2026). Recursive Language Models. arXiv:2512.24601.