Derivation Faithfulness of Compressed Reasoning Traces

Reasoning models now produce thinking traces that run to thousands of tokens, and nobody reads them whole. What people and downstream systems actually consume is a compressed version: a trace cut to a fifth of its length, a distilled list of "core steps," or a short summary of the thinking. Every existing check on these compressions asks the same question: is the final answer still right? This proposal asks a different one. When a long derivation is compressed, does the short version describe how the answer was actually reached, or does it invent a cleaner justification that the original reasoning never contained? A compressed trace can list only true statements and still misreport which step led to which. That is a form of hallucination about the reasoning itself, and current summary-faithfulness metrics are not designed to catch it.

We propose a measurement instrument and a study. The instrument is a structural faithfulness check. It turns the compressed trace into a small graph of claims, where each link records "this was derived from that." It then asks, link by link, whether the original trace contains a derivation connecting the two. A small set of human labels corrects the automated judge statistically, so that reported rates come with valid confidence intervals. The study applies the instrument across compression methods and compression ratios on math, multi-hop factual QA, and narrative soft-reasoning tasks. It tests whether answer accuracy survives compression while the stated derivation quietly stops matching the real one.

Core Claim

Compression that is optimized for final-answer accuracy can keep answers correct while making the stated reasoning unfaithful to the source trace. We expect the unfaithfulness to take two main shapes. The first is fabricated dependencies: the summary says B follows from A, but the source reached B another way. The second is concealed revision: the source tried a wrong path, backtracked, or hedged, and the summary presents a single confident line. Checking atomic facts alone cannot detect either, because every individual statement may appear in the source. Detecting them requires checking the dependency structure of the summary against the dependency structure of the source.

This separates three properties that the literature tends to merge. Answer correctness is what Extra-CoT and Breadcrumbs Reasoning measure. Internal validity asks whether the compressed trace is logically sound on its own terms, which is what FlowScore would measure if applied to it. Faithfulness to the source is what nobody measures for reasoning compression. A compressed trace can be valid yet unfaithful; this is post-hoc rationalization, and it is the case we care about.

Where This Sits in Prior Work

Several papers motivate the gap. Extra-CoT compresses mathematical chains of thought by up to 80% using a reinforcement-learning objective that penalizes wrong short answers. It reports accuracy only, so it cannot tell us whether its compressed traces still reflect the original derivation. Breadcrumbs Reasoning compresses the model's working memory rather than its text. It likewise reports task accuracy, and its reported arithmetic degradation on linear-system solving suggests that compression drops exact intermediate content. RACE documents the "right answer, wrong reason" failure in reasoning models. It also uses an 8B distiller to compress thinking traces into core steps, and it treats the fraction of discarded material as a hallucination signal. RACE therefore trusts its distiller, which is exactly the component whose fidelity we propose to audit. LEDGERMIND names phantom grounding (citing a source while fabricating the claim) and shows that free-form repair injects new errors. Our fabricated-dependency category is the reasoning-trace analogue of phantom grounding.

Other papers supply methods. ACUEval splits a summary into minimal atomic facts and verifies each against the source; its authors acknowledge that it ignores dependencies across units. We reuse its decomposition step and add the dependency check it leaves out. QAFactEval scores consistency by generating questions from the summary and answering them against the source. It is a strong atomic-level competitor. FlowScore prompts models to emit explicit conditions, steps, and cited parent steps as a graph, then verifies each step locally. We borrow its graph representation, but we verify links against the source trace instead of against local soundness.

For the statistics, Efficient Inference for Noisy LLM-as-a-Judge Evaluation shows that uncorrected automated judges give biased estimates and undercover the truth. It also shows that prediction-powered inference (PPI++), which combines many cheap model judgments with a few human labels, yields intervals that are "35-55% narrower than standard PPI." We apply PPI++ to our per-link judgments.

One paper constrains the design. FlowScore reports that its local verifier falsely flags steps that rely on unstated commonsense premises. This matters most on narrative reasoning such as MuSR. We therefore allow a link to count as supported through an implicit premise, provided the source trace itself invokes that premise, and we report narrative tasks separately.

Research Program

The metric. Take a source trace $T$ and a compressed version $S$. We first extract from $S$ a set of claims, each with the claims it presents as its premises, giving links $(i \to j)$. For every link, a judge model reads the relevant spans of $T$. The spans are retrieved by claim matching, so the judge never sees the whole trace at once, which avoids the long-context failures that CLIPPER documents. The judge assigns each link one of four labels: supported (T derives $j$ using $i$), rerouted (T asserts $j$ but reaches it through other premises), absent ($j$ is not in T), or overstated (T reaches $j$ tentatively, or after abandoning a contradicting branch, and $S$ drops that qualification). The last label operationalizes the epistemic overconfidence failure reported by Tang et al.. The headline score is the fraction of links labeled supported. We also report an atomic-only score computed on the same claims, so the gap between the two scores directly measures what dependency checking adds.

The test sets. We need two kinds of data. The first is a controlled set with known ground truth. We generate long traces with open reasoning models on GSM8K, MATH-500, HotpotQA, and MuSR, and write faithful compressions of them. We then apply targeted edits. Link rewiring gives a claim a different premise that also appears in the trace. Branch concealment removes a backtrack while keeping the conclusion. Hedge removal deletes a qualification. Answer-preserving fabrication inserts a plausible step that is not in the trace. All of these edits keep every atomic statement present in the source, so atomic checkers are tested fairly on the case they are claimed to miss.

The second kind is a natural set. It consists of real compressions from four compressors: Extra-CoT-style ratio-conditioned compression at several ratios, general token pruning in the LLMLingua style, the RACE-style distiller, and a plain "summarize your reasoning" prompt. Roughly 500 links from the natural set will be labeled by trained annotators. MuSR's authors found crowdsourcing inadequate, so we plan for specialists. These labels anchor the PPI++ correction.

The study. With a calibrated instrument, we measure faithfulness against compression ratio for each compressor and domain, alongside answer accuracy. The central question is whether the two curves come apart.

Baselines

The naive reference point is answer agreement: a compression counts as faithful if it reaches the same final answer as the source trace. This is effectively the criterion implied by accuracy-only evaluation in Extra-CoT and Breadcrumbs. It should score every item in our controlled set as faithful, because every edit preserves the answer. It anchors how much is missed by doing nothing.

The strongest existing competitor for detection is ACUEval, used with the trace as the source document and the compressed trace as the summary. It is the most direct fine-grained faithfulness method in the file, and it was designed for the same source-versus-summary comparison. Its reported result is a "3% improvement in balanced accuracy over state-of-the-art metrics" on LLMSUMMEVAL. We found no absolute score for it and no reported result on Lacuna for any metric on compressed reasoning traces. QAFactEval is a second atomic-level baseline, with a reported "77.5 balanced accuracy" on SummaC and 79.5 for its NLI variant. A holistic judge is the third: an evaluator model given the full trace and the compression with a faithfulness rubric in the BiGGen style, whose best reported human correlation is 0.623 Pearson for GPT-4-Turbo on BiGGen Bench.

Two further baselines target the validity-versus-faithfulness distinction. FlowScore is applied to the compressed trace alone; it reports 76.59% average precision on MATH and 83.08% on HumanEval for detecting reasoning errors, but it has no access to the source. RACE's discarded-content ratio serves as a cheap source-aware signal; RACE reports 77.62 AUROC on HotpotQA against 74.50 for SINdex, though for a different target, answer hallucination.

For fairness, every learned-judge baseline and our method use the same judge backbone. The judge is an open model, such as the StableBeluga-class model used in ACUEval, plus one frontier API model. All methods get the same retrieval over the source trace. Prompt tuning is limited to a matched budget on a shared development split. We report token cost per item, because our method, like ACUEval, pays for many small judgments.

Two ablations isolate the key ingredient. The first removes links and scores only claims, which should recover ACUEval-like behavior. The second keeps links but verifies them only for internal soundness, which should recover FlowScore-like behavior.

We count the method as beating the baselines only under three conditions. It must exceed the best baseline's balanced accuracy on the fabricated-dependency and concealed-revision edits by a margin outside the bootstrap interval. It must match that baseline on plain factual-error edits. And it must agree more closely with human labels on the natural set.

Why not just…?

Why not just use ACUEval or QAFactEval on the compressed trace? This is the real threat. If natural compressions rarely fabricate dependencies, and their unfaithfulness mostly shows up as atomic errors, atomic checkers will capture most of the signal at lower cost. ACUEval's own authors note that it cannot see cross-unit dependencies, which supports our concern but does not settle it. The link-free ablation settles it directly. On the controlled set we expect a large gap by construction. The informative number is the gap on the natural set. If it is small, we will report that atomic checking suffices for reasoning compression, which is itself a useful negative result.

Why not just ask a strong judge to compare the whole trace with its summary? Holistic judges miss subtle intermediate errors and suffer position bias on linear traces, according to FlowScore's analysis. CLIPPER shows that direct prompting over long inputs yields 73.1% erroneous claims, against 16.7% with structured compression. Source traces here run to thousands of tokens, so we expect holistic judging to degrade as trace length grows. We test this by stratifying all results by trace length.

Why not just check whether the compressed trace is valid, using FlowScore? Validity is not faithfulness. A cleaned-up rationalization can be perfectly sound while misreporting the derivation, and branch-concealment edits are designed to make traces more valid. The source-free ablation should score those edits as good. If it does not, our distinction is weaker than claimed.

Why not just keep the full trace and skip compression? People and pipelines demonstrably do not consume full traces; that is why Extra-CoT, Breadcrumbs, and RACE's distiller exist. The question is not whether to compress but whether the compressions can be trusted. That question needs an answer whether or not one approves of compression.

Proposed Experiments

The first decisive check is detection on the controlled set. We report balanced accuracy and AUROC per edit type for our metric, both ablations, and all baselines, using matched judge backbones. Success means the full metric catches rewired and concealed-branch edits that the atomic-only ablation scores as faithful, with no loss on ordinary factual edits.

The second is agreement on the natural set. We compare each method's link-level labels with human labels and report the PPI++-corrected population rate of unfaithful links with 95% intervals. The naive, uncorrected judge estimate is reported alongside, to show how much the correction moves the conclusion.

The third, and the substantive finding, is the ratio sweep. For Extra-CoT-style compression at ratios from mild to 0.2, and for the other compressors at matched lengths, we plot answer accuracy and corrected link faithfulness on GSM8K, MATH-500, HotpotQA, and MuSR. The proposal's hypothesis is that faithfulness declines well before accuracy does. We further expect answer-rewarded compressors to show more rerouted and overstated links, and token pruning to show more absent links. If faithfulness tracks accuracy closely on every task, the hypothesis is refuted, and accuracy-only evaluation of reasoning compression is vindicated.

The final check audits RACE's distiller as a special case. Hallucination detectors that rely on distilled traces are only as good as the distillation. We test whether RACE's AUROC on HotpotQA changes when its distilled steps are replaced by versions filtered to supported links only.