Claim Smoothing: When LLM Research Agents Turn 'Maybe' Into 'Definitely' — and How to Quantify the Overconfidence

A chilling scenario

Imagine hiring a research assistant to write a literature review. The report reads flawlessly: "Technique X improves efficiency by 40%", "Method Y outperforms baselines in all scenarios", "Researchers in field Z unanimously agree…". Then you check the original papers: the 40% gain holds only on one specific dataset; Method Y wins on three tasks and loses on five; "unanimous agreement" is actually two papers against three. The assistant fabricated nothing—every citation exists, every number checks out. But it did something sneakier: it turned 'maybe' into 'definitely', 'partial' into 'total', 'contested' into 'settled'.

This is the smoothing failure mode of LLM research agents, and the paper is the first to systematically quantify it.

What is smoothing?

Smoothing is when an LLM, synthesizing multiple documents, flattens diverse, conflicting, conditional evidence into a single decisive narrative. The paper describes real evidence as "evidentially jagged": some sources support a conclusion, some oppose it; some findings hold unconditionally, others only under specific conditions. Smoothing grinds away the jaggedness.

This differs from hallucination. Hallucinations invent non-existent facts and are relatively easy to catch via fact-checking. Smoothing is sneakier: every fact has a real source, but the conclusion is stronger than the source supports. You can't rebut it with "that citation is fake"—the citation is real, just over-interpreted.

SmoothScore: scoring overconfidence

The core contribution is SmoothScore, a claim-level scoring pipeline:

1. Extract claims — pull key conclusive assertions from the generated report. 2. Collect evidence — for each claim, search the source corpus for *all* topically relevant evidence, not just supporting evidence. 3. Cluster by aspects — group evidence into high-level aspects (e.g., performance on dataset A, dataset B, theoretical analysis, empirical limitations). 4. Score 1–5 per claim:

• 1: no smoothing — the claim faithfully reflects evidence complexity and conditionality.
• 2: slight smoothing — mild over-assertion, mostly faithful.
• 3: substantive smoothing — claim strength exceeds evidence; important counterexamples or conditions ignored.
• 4: severe smoothing — claim clearly conflicts with evidence or ignores key opposing evidence.
• 5: fundamentally misleading — the claim distorts the overall evidential direction.

Scoring uses an LLM judge with human validation: Spearman correlation between two annotators was 0.72, and between judge and annotators 0.69 — quite reliable for such a subjective task.

The numbers are striking

Testing 7 systems across 4 datasets:

• Gemini Deep Research is the worst smoother: 85–93% of claims score ≥3 (substantive), 22–35% score ≥4 (severe). Average SmoothScore 3.05–3.27.
• Open-source systems fare no better: DR Tulu and QUEST have severe-smoothing rates (≥4) of 7–23%, 13–26× higher than Luna and Sol. QUEST often produces output that "reads like a summary rather than a synthesis" — faithful to specific numbers but over-asserting in scope.
• Even the best systems smooth at least 22% of claims (Luna and Sol, scoring ≥3). No system gets below 20%.
• "Reads like a synthesis" ≠ "is a synthesis": Gemini produced the longest reports with the most citations and the most professional-sounding prose — and the most severe smoothing. Fluent, confident reports feel better; hedged ones feel rough. Smoothing is systematically rewarded stylistically.

Where smoothing comes from

A revealing analysis asked: do models actually read the documents? Mostly, no. On ResearcherBench:

| System | Cited docs | Actually opened | Open rate | Reports with no opened docs | |--------|-----------|-----------------|-----------|------------------------------| | Luna | 496 | 95 | 19% | 45/65 (69%) | | Sonnet 5 | 1,458 | 34 | 2% | 44/65 (68%) |

Sonnet 5 cited 1,458 documents but opened only 34 — a 2% open rate. Models are "pretending to synthesize", assembling plausible reports from search snippets, titles, and fragments without deep reading. Snippet-level evidence lacks the detail needed to preserve the "jaggedness".

Mitigations

• Provide full documents instead of search snippets: Luna's average SmoothScore dropped from 1.95 to 1.80; Sonnet 5's from 2.52 to 1.78. Substantive smoothing (≥3) fell from 29% to 12% (Luna) and 61% to 25% (Sonnet 5); severe smoothing (≥4) nearly disappeared.
• Add calibration instructions ("don't over-assert; state conflicts and conditions"): modest effect alone, but the largest gains come from combining full documents + calibration.

Key interaction: calibration instructions only work when the model has sufficient evidence. With only snippets, the model doesn't know what it's missing and can't mark uncertainties. Design implication: fix evidence access before fixing evidence expression.

Why this matters

Smoothing is隐蔽... (translating faithfully:) Smoothing is stealthier than hallucination — no fabricated facts, just systematic omission of hedges, which individually look like editorial polish but together constitute systematic misdirection. There is also a self-reinforcing tendency: smoothed reports enter training data, teaching future models that "this is how synthesis is written". SmoothScore's value is measurement: comparing systems, tracking improvement, setting deployment thresholds. If your company's Deep Research system averaged 3.27, you probably don't want an "auto-publish to website" button on it.

Honest limitations

• SmoothScore relies on an LLM judge, which may carry its own biases; the 73-sample human validation set is modest.
• The boundary between "smoothing" and legitimate "concision" is sometimes blurry — good synthesis does compress and simplify. The 1–5 scale quantifies, but edge cases remain subjective.
• Only 4 English datasets were tested; other languages and domains (legal analysis, market research) need further validation.

Closing thought

The counterintuitive insight: an LLM research agent's greatest danger isn't ignorance, it's pretending to know. It doesn't invent facts; it compresses facts into a cleaner shape than reality. That compression is a virtue in summarization — and a trap in research synthesis.

Next time an AI research report reads flawlessly, ask: were those "definitely"s actually "maybe"s in the source? Were those "all"s actually "some"s? Was that "consensus" actually two papers? Jaggedness is the true shape of evidence. Smooth narratives are often sanded-down truth.

---

Paper: Quantifying Claim Smoothing by LLM Research Agents Authors: Tiasa Singha Roy, Manya Wadhwa, Greg Durrett (New York University) arXiv: 2610.35793 Code: github.com/tiasa2/SmoothScore