Razor Burn: Why LLMs Nick Themselves on Induction and Abduction

Diagnosis is where AI systems start to look clever, then suddenly start charging consultancy rates.

Give a model a handful of symptoms, incident logs, customer complaints, or audit traces, and ask it what explains them. It will usually produce something plausible. Sometimes several plausible things. Occasionally an entire decorative shrubbery of plausible things. The practical question is not whether the model can invent an explanation. That bar is underground. The harder question is whether it can find the simplest explanation that accounts for the evidence without adding unnecessary machinery.

That is the point of Yunxin Sun and Abulhair Saparov’s paper, Do Language Models Follow Occam’s Razor? An Evaluation of Parsimony in Inductive and Abductive Reasoning.¹ The paper introduces InAbHyD, a synthetic benchmark for inductive and abductive hypothesis discovery. Its central move is useful because it does not treat “reasoning” as one undifferentiated glow around a model. It asks whether a model can infer hidden rules, missing memberships, and missing subtype relations from observations inside a controlled world model—and whether it can do so parsimoniously.

The misconception this paper quietly dismantles is familiar: if a model performs well on ordinary reasoning benchmarks, it should be able to handle open-ended hypothesis discovery. Not quite. Deduction asks the model to follow given rules toward a conclusion. Induction and abduction ask it to infer the rules or missing facts that would make the observations make sense. That is where the razor comes out. And, as the title suggests, the models nick themselves.

InAbHyD hides pieces of a world and asks the model to rebuild them

The benchmark’s mechanism matters more than its leaderboard. InAbHyD builds fictional first-order logic worlds as ontology trees. Each node is a concept. Nodes can have members and properties. Edges encode subtype relations. The complete world might contain facts such as “Amy is a wumpus,” “all wumpuses are bright,” or “every zorb is a wumpus,” except the concepts are fictional, precisely to avoid relying on memorised real-world associations.

Then the generator hides some axioms. Those hidden axioms become the ground-truth hypotheses. The benchmark produces observations that can be explained if the hidden axioms are recovered. The model sees the incomplete world model and the observations, written in natural language, and must propose hypotheses that explain all observations.

The generation pipeline is deliberately mechanical:

That last detail is easy to underestimate. The paper is not asking models to solve raw symbolic logic. It converts logical forms into natural language, then paraphrases them into more natural text. The authors also manually checked 200 paraphrased questions across difficulty levels and report that the paraphrase preserved the intended semantics. This is not a benchmark of whether a model can parse Prolog cosplay. It is a benchmark of whether a language model can infer the missing explanatory structure from text.

The benchmark includes three core hypothesis types:

That combination is the benchmark’s first contribution. It gives researchers a way to vary the kind of missing explanation, the height of the ontology tree, and whether one or multiple hypotheses are needed. The dataset contains over 2,000 reasoning questions, and the generation code can in principle produce many more.

The business translation is simple: this is the kind of task hiding inside root-cause analysis, fraud triage, compliance review, support escalation, medical reasoning, and incident diagnosis. The benchmark is synthetic, so it is not a production proxy. But the pattern is highly relevant. Many enterprise agents will not merely retrieve answers. They will infer missing links from partial evidence. That is where valid-but-bloated explanations become operationally expensive.

Weak correctness is not the same as useful explanation

The paper’s most important evaluation choice is the separation of three metrics.

Strong accuracy means the model’s hypotheses exactly match the ground-truth simplest hypotheses. Weak accuracy means the hypotheses can explain all observations, even if they are not the simplest. Hypothesis quality measures how well the answer follows Occam’s Razor.

This distinction is the article’s hinge. A weakly correct answer can still be bad. It can explain every observation by adding redundant rules, restating observations as hypotheses, or proposing overly specific explanations when a general one is available. That is not harmless. In a business workflow, a model that explains all anomalies by inventing five local exceptions instead of one systemic cause has technically “explained” the evidence. It has also made the organisation stupider, only with confidence.

The quality metric rewards hypotheses that are reused across proof trees and penalises unnecessary hypotheses. For valid explanations, the paper defines quality as the ratio between average use of the candidate hypotheses and average use of the ground-truth hypotheses:

If the candidate hypotheses cannot explain the observations, the score is zero. The ground-truth hypothesis set scores one.

The intuition is cleaner than the notation. A good explanation should do work. If one general hypothesis explains many observations, it is usually better than several narrow hypotheses that each explain one observation. The paper further validates this metric with a human study: three graduate-level STEM annotators evaluated 100 reasoning questions, each with three candidate hypotheses. The metric agreed with the human-preferred hypothesis 79% of the time, compared with 33.3% for random choice and 51% for a length-based heuristic. The annotators’ agreement was substantial, with Fleiss’ $\kappa = 0.75$.

That validation is not a universal theory of explanation quality. It is a reasonable sanity check that the metric is measuring something closer to parsimony than token count. Occam’s Razor, disappointingly for the lazy, is not just “write less.”

The easy setting flatters models; the coupled setting exposes them

The experimental design has a useful sequence. First, the authors test single-hypothesis cases. Then they test multiple-hypothesis cases. Then they test whether in-context learning, reasoning models, and reasoning-oriented prompting help. The purposes are distinct:

The single-hypothesis results are encouraging at the shallow end. Under a height-1 ontology tree, all tested models achieve above 80% weak and strong accuracy across the three task types. As tree height increases, performance generally falls. The paper notes one interesting asymmetry: models perform better on membership inference than on property or subtype inference. For example, Gemma3-27B reaches near 50% weak accuracy on height-4 membership inference, while dropping to around 10% on height-4 subtype inference.

That makes sense mechanically. Membership inference can sometimes be solved by searching for a concept whose properties match the observed entity. Subtype and property inference demand more navigation through the ontology. Search is not the same as structural reasoning, although product demos often dress them in the same suit.

The multi-hypothesis setting is where the benchmark becomes more revealing. The paper reports that when ontology height increases from 1 to 2, accuracy for all models except GPT-4o drops from above 80% to below 50%. This is striking because the average number of ground-truth hypotheses rises only from 3.0 to 3.5. The average number of observations also grows linearly. The number of world-model axioms, however, grows much faster: in the multi-hypothesis examples, average world-model axioms rise from 9.0 at height 1 to 46.8 at height 4, while observations rise from 10.0 to 20.0 and ground-truth hypotheses from 3.0 to 6.6.

The problem is not just “more facts.” It is coupling. Multiple missing explanations interact with a deeper ontology. The model has to decide which general rule, membership, or subtype relation best compresses the observations. It must avoid explaining local symptoms one by one when a higher-level explanation is available. This is exactly the kind of thing that looks easy in a slide deck and becomes irritating in production.

At height 4, the paper reports that all models still achieve at least 20% weak accuracy, but strong accuracy and quality are extremely low. That gap is the result to remember. The model can often produce something that covers the observations. It fails to produce the simplest correct explanatory structure.

Reasoning tricks help, but they do not restore the razor

The authors test two families of improvement: in-context demonstrations and reasoning-enhanced prompting or models.

For in-context learning, they use eight demonstrations with ground-truth hypotheses and chain-of-thought proofs. They compare in-distribution demonstrations, where examples match the test question’s ontology height, with out-of-distribution demonstrations, where each demonstration is a single-hypothesis height-1 example. In-distribution demonstrations are slightly more helpful, especially for strong accuracy and quality at heights 3 and 4. Out-of-distribution demonstrations do not produce a significant improvement.

This is a useful practical result. Examples help when they teach the model the shape of the actual task. Toy examples do not necessarily transfer to more structured cases. Shocking, I know: showing someone a tricycle does not qualify them to land a cargo aircraft.

The paper also compares Llama3-70B with larger reasoning models, GPT-5.4 and o3, under chain-of-thought and buffer-of-thought prompting. The stronger models consistently outperform Llama3-70B across weak accuracy, strong accuracy, and quality. Buffer-of-thought generally improves over chain-of-thought, especially for weak accuracy. But the gap between weak and strong accuracy persists. Lower-height cases can approach near-perfect weak accuracy while strong accuracy remains lower and degrades faster with height.

That means reasoning models and prompting improve coverage more readily than parsimony. They make the model better at finding explanations, but not reliably better at finding the right level of explanation. For enterprise AI, that difference is not academic. A system that covers every anomaly but cannot compress them into the minimal causal story will bury reviewers under technically defensible noise.

The failure modes are mostly failures of structure, not vocabulary

The paper’s error analysis is especially useful because it shows what “low quality” looks like. The authors manually inspect 200 model responses and then use a GPT-4o judge with in-context examples to categorise error patterns across responses. They identify four recurring failures:

The most significant error across ontology heights is ignoring the ontology and producing unnecessary hypotheses. That is precisely an Occam failure. The model is not always failing because it cannot find any valid explanation. It fails because it does not respect the explanatory economy of the world model it has been given.

This matters for agent design. If an enterprise agent has access to a product taxonomy, control framework, system dependency graph, or organisational hierarchy, it should use that structure to compress hypotheses. If it treats every observation as a fresh island, the output becomes a pile of local explanations. It may look thorough. It is actually wasteful.

What businesses should borrow is the stress-test pattern

The paper does not show that InAbHyD predicts production performance. It explicitly frames the dataset as synthetic and out-of-distribution. The world models are first-order logic ontology trees. Real enterprises contain ambiguous documents, stale process maps, conflicting definitions, partial observability, incentives, exceptions, and people using the same word to mean three different things before lunch.

So the business relevance is not “run InAbHyD and choose a vendor.” The relevance is that InAbHyD shows how to design better internal evaluation tasks for agents that must infer missing causes or rules.

A useful business version would look like this:

The practical pathway is evaluation-first. Before trusting an agent to diagnose production incidents, compliance breaches, procurement anomalies, or customer churn, build a local InAbHyD-like test. Give it a controlled domain model. Hide known links. Provide observations. Score not only whether it explains the observations, but whether it finds the most economical explanation.

That score should not be merely “did the answer sound reasonable?” Reasonable is cheap. The test should ask:

1. Did the hypotheses explain all observations?
2. Did they match the known missing rule, cause, or relationship?
3. Did they add unnecessary explanations?
4. Did they reverse category or causal direction?
5. Did they introduce unsupported entities?
6. Did they use the existing ontology instead of ignoring it?

This is where ROI appears, if one insists on using the term. The value is not that the agent writes a prettier incident report. It is that it reduces false branches in investigation. It prevents analysts from chasing redundant causes. It makes escalation shorter because the explanation is compact enough to review.

The boundary is synthetic logic, not synthetic usefulness

There are two main limitations to keep in view.

First, InAbHyD is synthetic. That is a strength for controlled evaluation and a weakness for direct deployment claims. Synthetic fictional worlds reduce contamination and make ground truth available. They also omit the semantic mess that makes real-world diagnosis painful. A model that does well here has not proved it can handle a bank’s risk taxonomy or a hospital’s patient history. A model that fails here, however, has revealed something uncomfortable: even under clean conditions, parsimony is hard.

Second, the benchmark is based on first-order logic. FOL is expressive enough for ontology trees, membership, properties, subtype relations, and proof-based evaluation. It is not the full shape of real explanation. Many business explanations involve probabilistic causality, temporal sequences, feedback loops, counterfactuals, incentives, and missing data. The paper itself notes that higher-order logic could support more complex future benchmarks.

Those limitations do not weaken the main lesson. They locate it. InAbHyD is not a mirror of the enterprise. It is a wind tunnel. The plane still has to fly outside, but if the wing comes off in the tunnel, perhaps delay the champagne.

The next agent benchmark should grade explanatory restraint

The paper’s contribution is not just another benchmark with another set of bars. Its contribution is a sharper evaluation question: can a language model infer missing explanations and keep them simple?

That question is business-relevant because many valuable AI workflows are not answer lookup. They are hypothesis discovery. A support agent infers the likely cause of a customer’s issue. A fraud agent infers the pattern behind suspicious activity. A compliance agent infers whether a breach reflects one employee mistake or a control gap. A product analytics agent infers why a metric moved. In each case, many explanations can fit the observations. The useful one is usually the simplest explanation that respects the structure already known.

InAbHyD shows that current models can perform well in shallow, isolated settings, but degrade when the ontology deepens and hypotheses multiply. It also shows that prompting and stronger reasoning models help without removing the weak-versus-strong gap. The models can often cover the evidence. They struggle to shave the explanation down to the right shape.

That is the razor burn. The model does not always fail by saying something absurd. It often fails by saying too much, at the wrong level, with just enough plausibility to keep a human busy.

For enterprise AI, the design implication is unfashionably practical: build agents with ontology grounding, redundancy checks, directionality checks, and parsimony-aware review. Do not merely ask whether the agent can explain. Ask whether it can stop explaining once the simplest explanation has done the job.

Cognaptus: Automate the Present, Incubate the Future.