Papers on entangled photon imaging rarely state their key results as plain numbers sitting in a table. They state derived quantities. A sentence such as "inter-pixel correlations double the sampling resolution" rests on a 32 µm physical pixel and a 16 µm effective pixel. "A 1.7× broader modulation transfer function" is a ratio of two measured curves. "10–20% resolution gain for three- and four-photon states" follows from a narrowing factor of $\sqrt{2(n-1)/n}$ that appears in the paper's prose, not in its tables. Verifying such a sentence means finding the right table cells, converting their units, and applying a relation the paper itself states. Today's table fact-checkers are trained to spot claims that are entailed or contradicted by cells. They handle this kind of claim poorly, and they handle worst the case where the needed quantity simply is not present.
This note proposes a verifier for quantitative claims about results tables. It turns each claim into a small program in which every number must be traced to its source: a table cell (with its unit read from the header), or a relation retrieved from the paper's text. If the program runs and every number has a source, the verdict is supported or refuted. If any number has no source, the claim is labeled "unknown." The cells the program used are the evidence. Entangled photon imaging serves as the test domain because its claims are dense with ratios, scaling laws, and unit conversions. The method itself is not physics-specific.
Core Claim
The claim is narrow and checkable. On scientific tables, errors concentrate in three places: claims whose truth depends on a quantity derived from cells rather than read off them, claims that need a unit conversion, and claims that should be labeled unknown. A verifier that requires a source for every number fixes all three. It does so because "can every number in this claim be traced?" is the same question as "is this claim verifiable?" It also makes evidence-cell selection a byproduct of verification rather than a separate classifier.
The research file motivates the gap from two sides. First, the SemEval-2021 scientific-table shared task, SEM-TAB-FACTS, built three-way verification plus cell-evidence selection over scientific tables. Its best evidence-selection system reached only 65.17% F1, far below the best verdict systems. The organizers also report that refuted and unknown statements are harder than entailed ones. Second, the TabFact dataset paper, TabFact, introduced the simple/complex split and showed early models falling to about 58% on complex multi-row claims.
The method borrows from program-based verifiers. TabFact's Latent Program Algorithm ranks candidate programs and executes the best one. ProgVGAT turns executed programs back into sentences and feeds them to a graph network. Binder lets a language model write SQL or Python calls. None of these require every literal in a program to have a recorded source, none reason about units, and none treat a missing source as the signal for "unknown."
PASTA constrains the design. It shows that symbolic programs can be "spurious," reaching the right label through wrong logic. Our source requirement is partly a guard against this: a program that lands on the right answer using an untraced constant is rejected.
I considered and rejected the more literal bridge between the two fields. That bridge would treat cell-pair attention as an analogue of the photon joint probability distribution in Pixel super-resolution with spatially entangled photons, or treat cell pruning as the "lost photon" conditioning of Lost photon enhances superresolution. Both reduce to pairwise attention and learned masking, which structure-aware encoders such as Table Fact Verification with Structure-Aware Transformer and LKA already provide. The physics papers are therefore used for what they actually offer: a family of real, checkable quantitative claims.
Research Program
The first deliverable is a diagnostic set, provisionally called QImg-Claims. It is drawn from tables and results paragraphs of imaging-physics papers: entangled-photon work of the kind in the file, plus classical imaging papers to avoid a single-lab style. The quantities include pixel pitch, modulation transfer function width, point-spread-function width, photon number, acquisition time, and the dimensionality of the measured entanglement. Typical sources are the camera studies in Imaging and certifying high-dimensional entanglement with a single-photon avalanche diode camera and Towards real-time quantum imaging with single photon avalanche diode cameras.
Annotators follow the SEM-TAB-FACTS protocol. They write entailed, refuted, and unknown statements and mark the evidence cells. Each statement is also tagged by type: direct lookup, arithmetic over cells, unit conversion, or derivation through a relation stated in the text. The target is roughly 300 tables and 1,500 statements. This is a target, not a measured quantity; the real count depends on how many papers have usable tables.
The verifier has three parts:
- Program proposer. A language model writes the program in a small typed language. The language has the PASTA-style operations of filter, aggregate, superlative, comparison, ordinal, and unique, extended with unit-bearing arithmetic and a call that applies a retrieved textual relation.
- Executor. The executor carries units through every operation. It sets numeric tolerance from the significant figures of the source cells.
- Decision rule. If any literal in the program lacks a source, the verdict is unknown. Otherwise, the verdict is the executed comparison. The evidence prediction is the set of bound cells.
We sample several programs and decide by agreement among those that pass the source check. This reduces the effect of single spurious programs.
Baselines
The naive reference points are the majority class and a claim-only classifier. The file reports claim-only accuracy of 0.58 on FEVEROUS (FEVEROUS), a reminder that wording alone can leak labels. We will measure the same bias on QImg-Claims, and any method must clearly beat it.
The strongest trained table model is PASTA. It reports 90.6% accuracy on TabFact-small, 85.6% on TabFact-complex, and a gain of more than 5 points over all baselines on the SEM-TAB-FACTS scientific tables (PASTA). We will fine-tune it on the SEM-TAB-FACTS training split and evaluate it on QImg-Claims with the same splits.
For the scientific-table setting itself, the top shared-task numbers are the bar to clear:
- Verdicts: King001 reached 84.48% three-way F1.
- Evidence selection: BreakingBERT@IITK reached 65.17% F1.
Both are reported in SEM-TAB-FACTS.
The most direct competitor for our approach is Binder, a language model writing SQL or Python. No reported result for Binder on SEM-TAB-FACTS or TabFact was found on Lacuna. We will therefore run it ourselves with the same language model as our proposer, the same number of sampled programs, and the same prompt-tuning budget.
Two weaker program-free references complete the set. One is direct prompting of a large language model; the file reports ChatGPT at about 74–76% on TabFact-small (Are Large Language Models Table-based Fact-Checkers?). The other is that model with step-by-step reasoning and a calculator tool.
Three ablations isolate the key ingredients, each removed separately from the full system:
- drop unit handling;
- drop the source-based unknown rule and let the model emit unknown freely;
- drop retrieval of relations from the paper's text.
"Beating" a baseline means a statistically significant gain in three-way macro F1 on QImg-Claims, a gain in unknown-class F1, and evidence F1 above 65.17%, all with no loss on SEM-TAB-FACTS Task A.
Why not just…?
Why not just fine-tune PASTA on scientific tables? PASTA already generalizes to SEM-TAB-FACTS. However, it reads a linearized, pruned table and has no way to apply a formula that appears only in the paper's prose or to convert micrometres to pixels. We expect it to stay competitive on lookup and arithmetic claims and to fall behind only on unit-conversion and textual-relation claims. The per-type breakdown will show whether that expectation holds.
Why not just use Binder with a Python executor? This is the real threat. Python can compute $\sqrt{2(n-1)/n}$ as easily as our executor can. If Binder, given the same retrieved text, matches us on every claim type, then the source-tracing rule adds nothing. In that case the contribution shrinks to the diagnostic set plus a finding about where scientific verifiers fail. Ablation 2 against Binder is the test that settles this. Our prediction is that Binder fills missing quantities with plausible constants and so misses the unknown class.
Why not just prompt a stronger language model with step-by-step reasoning? The file reports that prompted models plateau around 75% on TabFact-small, that more in-context examples hurt, and that arithmetic over many rows is where they invent numbers (Are Large Language Models Table-based Fact-Checkers?). Those results come from an older model generation. We will rerun this baseline with the same model we use as our program proposer. That isolates the effect of the program structure from the effect of model strength.
Proposed Experiments
The decisive check is a per-type, per-class comparison on QImg-Claims. Our hypothesis predicts a specific pattern: gains over PASTA and Binder concentrated in unit-conversion and textual-relation claims and in the unknown class, with parity on lookups. A uniform gain across all types would instead suggest that the language model doing the proposing, not the source rule, explains the result.
The second check is evidence selection on SEM-TAB-FACTS Task B. Here we test whether cells bound by programs beat the 65.17% F1 obtained from a dedicated classifier.
The third check is TabFact-complex, to confirm the typed language costs nothing on general Wikipedia tables relative to PASTA's 85.6%.
Finally, we audit spurious programs. Annotators label 200 correctly-decided programs as right or wrong in their reasoning, under each system. This tests whether the source requirement lowers the spurious rate that PASTA identified as a weakness of symbolic verifiers.
References
- SemEval-2021 Task 9: Fact Verification and Evidence Finding for Tabular Data in Scientific Documents (SEM-TAB-FACTS)
- TabFact: A Large-scale Dataset for Table-based Fact Verification
- Program Enhanced Fact Verification with Verbalization and Graph Attention Network
- PASTA: Table-Operations Aware Fact Verification via Sentence-Table Cloze Pre-training
- Pixel super-resolution with spatially entangled photons
- Lost photon enhances superresolution
- Table Fact Verification with Structure-Aware Transformer
- Table-based Fact Verification with Self-labeled Keypoint Alignment
- Imaging and certifying high-dimensional entanglement with a single-photon avalanche diode camera
- Towards real-time quantum imaging with single photon avalanche diode cameras
- FEVEROUS: Fact Extraction and VERification Over Unstructured and Structured information
- Are Large Language Models Table-based Fact-Checkers?