A Four-Stage Decomposition of Word-Problem Solving and Mechanistic Fragility in LLM Math Reasoning
cs.AI · cs.LG
Quick verdict
Claims a four-stage internal pipeline for LLM word-problem solving and localizes distractor fragility to Operation Planning heads; strong, well-controlled multi-method evidence with interpretive leaps and model-specific caveats.
Summary
The paper unifies two opposing findings on LLM math word problems - behavioral collapse under irrelevant 'NoOp' clauses and mechanistic evidence for staged abstraction - into one account. On Llama-3.3-70B it identifies four residual-stream stage boundaries: Schema Abstraction (L22), Operation Planning (L36), Operand Binding (L40), Computation (L80), via three converging methods: within/cross-template residual cosine dynamics, template-disjoint linear probes (entity decodability 0.78 to 0.53 at L22; operand numbers 0.77 at L40; final answer 0.93 at L80), and cross-prompt activation patching on padded-Symbolic/GSM-P1 pairs that differ in exactly one operation, which identifies L36 as the Operation Planning boundary. Filler-correct-to-NoOp-wrong patching collapses in the same L22-L36 band, localizing NoOp fragility there. An engagement-anchored Direct Logit Attribution selects ~297 attention heads split into anti- and pro-engagement sets: ablating them collapses Symbolic-correct accuracy to 0.02 versus >=0.98 for size-matched random heads, amplifying the anti-engagement set recovers 0.51 of NoOp-wrong cases (random 0.19), and blind annotation attributes 0.91 of induced failures to planning errors versus 0.10 for random heads. Replications cover CoT/pre-reasoning readouts, Qwen-2.5-14B, Gemma-2-9B, SVAMP, and a PhantomWiki proof of concept. Limitations: the dominant head subset is model-specific (H+ on Gemma, H- on Llama/Qwen), Operation Planning lacks any direct probe signature, head sets are selected and tested on overlapping data, and GSM-NoOp is a re-creation rather than the unreleased original.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- LLM computation on GSM-style word problems decomposes into four sequential stages (Schema Abstraction, Operation Planning, Operand Binding, Computation) with identifiable layer boundaries (L22/L36/L40/L80 in Llama-3.3-70B), established by convergent cosine, probe, and patching evidence.
- NoOp distractor-induced failure localizes to the Operation Planning stage (L22-L36 band), not to Schema Abstraction or Operand Binding, by mirrored question-span patching from Filler-correct into NoOp-wrong pairs.
- A compact set of ~297 attention heads in that band causally implements Operation Planning: ablation collapses correct solving to 0.02 versus >=0.98 for size-matched random heads, and 0.91 of induced failures are planning-type under blind annotation.
- Amplifying the anti-engagement subset alone recovers 0.51 of NoOp failures on Llama-3.3-70B (~4 sigma over the 0.19 random baseline), a training-free lever indicating where repair must act.
- The staged structure reproduces on Qwen-2.5-14B and Gemma-2-9B with model-specific boundaries and transfers at probe level to SVAMP and, as a proof of concept, to PhantomWiki multi-hop QA.
Strengths
Convergent multi-method identification of stage boundaries
Section 5 triangulates boundaries with methods sharing no parameters: residual cosine similarity under an automatic criterion (drop >=0.025 at z>=3), template-disjoint linear probes, and causal patching (commit-readiness rising 0.08 to 0.88 across L30-L36, Fig. 4b). The paper is explicit about which method establishes which stage, and that L36 has no probe signature.
Unusually careful control datasets
App. A and Table 4 introduce padded-Symbolic (length-matched to GSM-P1, accuracy within -0.004/-0.016), digit-matched GSM-Filler, and digit-free GSM-Filler-DF (statistically null), separating length, digit tokens, and clause semantics; App. C/Fig. 10 re-run the NoOp diagnosis on Filler-DF to rule out soft-NoOp artifacts.
Bidirectional causal validation with size-matched controls and blind labeling
Breadth of robustness checks and honest cross-model reporting
Weaknesses
Operation Planning is the least directly evidenced stage
Head-set selection and evaluation share data
Subset roles and rescue effects do not generalize as framed
Analyses condition on model-correct subsets in a weak direct-prompting regime, with inconsistent template accounting
Re-created NoOp and unreported annotation reliability
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What are the four stages of math word-problem solving in LLMs according to this paper?
The paper claims LLMs process grade-school word problems through a four-stage pipeline in the residual stream: (1) Schema Abstraction, where surface narrative words are dropped and the relational structure kept (around layer 22 of 80 in Llama-3.3-70B); (2) Operation Planning, where the sequence of arithmetic operations is chosen (layer 36); (3) Operand Binding, where concrete numbers from the problem are inserted into the plan (layer 40); and (4) Computation, where the final answer is produced (layer 80). Boundaries were found by residual cosine similarity, linear probes, and activation patching.
Why do LLMs fail when an irrelevant sentence is added to a math problem (the NoOp effect)?
By patching activations from a correctly solved problem with a neutral filler clause into a wrongly solved problem with an irrelevant clause, the authors show the corruption enters during the Operation Planning band (layers 22-36 in Llama-3.3-70B): the model locks in a wrong operation plan, and a set of attention heads that read the distractor clause carries the corruption into that plan. Once the plan crystallizes at layer 36, later processing no longer responds to repair of the question representation.
Can this research fix distracted LLM reasoning without retraining?
Only partially and not yet deployably. Doubling the output of a specific anti-engagement attention-head subset recovers 51% of NoOp-failed problems on Llama-3.3-70B, versus 19% from scaling size-matched random heads. The effect size varies by model (the effective subset differs on Gemma and Qwen), and the authors describe repair as future work; the finding mainly tells future repair methods where to act.
Which models were studied and do the findings generalize beyond one model?
The main analysis uses Llama-3.3-70B-Instruct; the four-stage structure and the NoOp diagnosis replicate on Gemma-2-9B-it and Qwen-2.5-14B-Instruct with model-specific layer boundaries, and probe-level signatures transfer to the SVAMP benchmark and to a PhantomWiki multi-hop question-answering proof of concept. However, which attention-head subset causes versus suppresses distraction flips between models.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures were extracted as text only; curve shapes and error bars could not be inspected directly, so figure-level judgments rely on embedded numeric labels and captions.
- The original GSM-NoOp dataset was never released; all NoOp findings concern the authors' Claude Opus 4.7-generated re-creation, which may differ in difficulty from the benchmark that motivated the study.
- Appendix experiments (Figs. 12-18, Tables 7-8) were verified through captions and reported numbers only.
- Template-count discrepancies across sections (50 vs 75 vs 87/99/100) could not be resolved from the text and are flagged as reviewer questions rather than confirmed errors.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.