Pith. sign in

REVIEW 14 cited by

Arithmetic Without Algorithms: Language Models Solve Math With a Bag of Heuristics

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2410.21272 v2 pith:ZO25V2XL submitted 2024-10-28 cs.CL

classification cs.CL
keywords arithmeticheuristicneuronsalgorithmsheuristicsllmsmechanismmodel
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

Do large language models (LLMs) solve reasoning tasks by learning robust generalizable algorithms, or do they memorize training data? To investigate this question, we use arithmetic reasoning as a representative task. Using causal analysis, we identify a subset of the model (a circuit) that explains most of the model's behavior for basic arithmetic logic and examine its functionality. By zooming in on the level of individual circuit neurons, we discover a sparse set of important neurons that implement simple heuristics. Each heuristic identifies a numerical input pattern and outputs corresponding answers. We hypothesize that the combination of these heuristic neurons is the mechanism used to produce correct arithmetic answers. To test this, we categorize each neuron into several heuristic types-such as neurons that activate when an operand falls within a certain range-and find that the unordered combination of these heuristic types is the mechanism that explains most of the model's accuracy on arithmetic prompts. Finally, we demonstrate that this mechanism appears as the main source of arithmetic accuracy early in training. Overall, our experimental results across several LLMs show that LLMs perform arithmetic using neither robust algorithms nor memorization; rather, they rely on a "bag of heuristics".

Discussion (0). Sign in to comment.

Forward citations

Cited by 14 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Can Transformers Really Do It All? On the Compatibility of Inductive Biases Across Tasks

    cs.LG 2026-07 conditional novelty 7.0 of 10

    Learned replacement non-linearities show transformers are rarely optimal for algorithmic tasks, with benefits that are task-specific, while language/code gains are smaller and more transferable.

  2. Neuron-Anchored Rule Extraction for Large Language Models via Contrastive Hierarchical Ablation

    cs.LG 2026-05 unverdicted novelty 7.0 of 10

    MechaRule localizes agonist neurons in LLMs via contrastive hierarchical ablation to ground rule extraction in circuitry, recalling 96.8% of high-effect neurons and reducing task performance when suppressed.

  3. Are Arithmetic Heuristic Neurons Form-Invariant? A Mechanistic Analysis of Symbols, Text, and Code in LLMs

    cs.CL 2026-07 conditional novelty 6.0 of 10

    Llama-3 arithmetic is computed by a shared neuron set across symbolic, text, and code formats, and cross-format failures are activation-state differences, not distinct circuits.

  4. ToxiREX: A Dataset on Toxic REasoning in ConteXt

    cs.CL 2026-06 unverdicted novelty 6.0 of 10

    ToxiREX is a new dataset of 128k Reddit comments in six languages with hierarchical annotations for implicit toxicity in conversational context based on an existing reasoning schema.

  5. Represented Is Not Computed: A Causal Test of Candidate Algorithmic Intermediates in a Transformer

    cs.LG 2026-05 unverdicted novelty 6.0 of 10

    Transformer represents but does not causally transmit staged algorithmic intermediates for base-digit extraction, diverging from probe predictions.

  6. Neuron-Anchored Rule Extraction for Large Language Models via Contrastive Hierarchical Ablation

    cs.LG 2026-05 unverdicted novelty 6.0 of 10

    MechaRule localizes sparse agonist neurons via contrastive hierarchical ablation and adaptive group testing to ground rule extraction, recalling 97% of high-effect activations at 2.14% cost while enabling near-total e...

  7. Mitigating Prompt-Induced Cognitive Biases in General-Purpose AI for Software Engineering

    cs.SE 2026-04 unverdicted novelty 6.0 of 10

    A prompting method that forces GPAI models to state SE best practices before deciding reduces prompt-induced cognitive biases by 51% on average across eight tested biases.

  8. Generalization in LLM Problem Solving: The Case of the Shortest Path

    cs.AI 2026-04 unverdicted novelty 6.0 of 10

    LLMs show strong spatial generalization to unseen maps in shortest-path tasks but fail length scaling due to recursive instability, with data coverage setting hard limits.

  9. The Model Says Walk: How Surface Heuristics Override Implicit Constraints in LLM Reasoning

    cs.CL 2026-03 conditional novelty 6.0 of 10

    LLMs systematically let surface heuristics override unstated constraints; a new 500-item benchmark quantifies this and shows goal-decomposition prompting partially mitigates it.

  10. The Model Says Walk: How Surface Heuristics Override Implicit Constraints in LLM Reasoning

    cs.CL 2026-03 unverdicted novelty 6.0 of 10

    LLMs prioritize surface heuristics such as distance cues over implicit constraints in reasoning tasks, with the new HOB benchmark showing no model exceeds 75% strict accuracy and hints recovering performance.

  11. All for One: LLMs Solve Mental Math at the Last Token With Information Transferred From Other Tokens

    cs.CL 2025-09 conditional novelty 6.0 of 10

    LLMs solve arithmetic in-context via an All-for-One pattern, with all input-specific computation occurring at the last token after a two-layer information transfer window.

  12. Modular Arithmetic: Language Models Solve Math Digit by Digit

    cs.CL 2025-08 conditional novelty 6.0 of 10

    LLMs perform 3-digit addition and subtraction via digit-position-specific MLP circuits that can be intervened upon to change individual output digits.

  13. Sparse probes and murky physics: a case study of interpretability challenges in a foundation model for continuum dynamics

    cs.LG 2026-06 unverdicted novelty 5.0 of 10

    Case study applies SAE probing with enstrophy triage to a continuum-dynamics foundation model and reports intermittent feature consistency that does not align with standard physics while linking some output discrepanc...

  14. Reward Modeling for Reinforcement Learning-Based LLM Reasoning: Design, Challenges, and Evaluation

    cs.LG 2026-02 conditional novelty 4.0 of 10

    A taxonomy-driven survey arguing that reward design is the central mechanism shaping reliable LLM reasoning, with maps of reward paradigms, reward-hacking failure modes, and benchmark pitfalls.

Pith tools