mlsanityData to 14 Oct 2026

Twin Worlds: Equivariance-Based Abstention for Evidence-Grounded Reasoning

arXiv 2608.28018 ↗First posted 28 Aug 2026cs.CL cs.AI cs.LGcited by 1

On the map: LLM reasoning → Reasoning benchmarks and evaluation

Abstract (arXiv)

Knowledge-intensive reasoning requires Large Language Models (LLMs) to ground answers in provided evidence. When evidence is insufficient, it is desirable that models abstain rather than confidently generating unsupported answers. Existing abstention methods rely on uncertainty estimation or evidence sufficiency checks, but neither tests whether the reasoning process for generation, driven by the interaction of provided evidence and the model's internal memory parameters, is actually grounded in the evidence. A key contributing factor is that entity mentions in context activate memorised associations, causing models to generate plausible responses ungrounded in evidence. We propose Twin Worlds (TW), a framework for improving reliability in knowledge-intensive reasoning through equivariance-based abstention: unlike invariance, which requires outputs to remain unchanged, equivariance requires outputs to transform correspondingly under entity substitutions. A model grounded in the evidence should produce answers that shift consistently when entities are substituted while their relations are preserved. TW constructs multiple worlds via typed substitutions of the original input that preserve relational structure while reducing parametric priors, and uses equivariance violations as an abstention signal. Across four benchmarks and three model backbones, TW identifies when answers are not reliably grounded in the provided evidence and outperforms uncertainty- and sufficiency-based baselines.

Key facts. Cited by 1 paper in the map (data through 14 Oct 2026). Placed in Reasoning benchmarks and evaluation, one of 17 directions in LLM reasoning. Builds on 48 of its 54 references that are in the map. First posted on arXiv 28 Aug 2026.

Builds on

48 of 54 references.

Closest ancestors · 4

Same direction, newest first

  1. 2025SynthWorlds: Controlled Parallel Worlds for Disentangling Reasoning and Knowledge in Language Models28 Oct
  2. 2025AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions10 Jun
  3. 2025Memorization vs. Reasoning: Updating LLMs with New Knowledge16 Apr
  4. 2022ReAct: Synergizing Reasoning and Acting in Language Models6 Oct

Other references · 44

  1. 2020Retrieval-Augmented Generation for Knowledge-Intensive NLP Taskscited by 9,297
  2. 2022Large Language Models are Zero-Shot Reasonerscited by 4,374
  3. 2018HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answeringcited by 3,754
  4. 2017TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehensioncited by 2,955
  5. 2023ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIscited by 1,947
  6. 2018FEVER: a large-scale dataset for Fact Extraction and VERificationcited by 1,375
  7. 2016Group Equivariant Convolutional Networkscited by 1,165
  8. 2022When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memoriescited by 1,128
  9. 2020Beyond Accuracy: Behavioral Testing of NLP models with CheckListcited by 1,006
  10. 2019Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inferencecited by 876
  11. 2020SE(3)-Transformers: 3D Roto-Translation Equivariant Attention Networkscited by 607
  12. 2023Dissecting Recall of Factual Associations in Auto-Regressive Language Modelscited by 528
32 more
  1. 2018On the Generalization of Equivariance and Convolution in Neural Networks to the Action of Compact Groupscited by 300
  2. 2019Negated and Misprimed Probes for Pretrained Language Models: Birds Can Talk, But Cannot Flycited by 243
  3. 2021A Practical Method for Constructing Equivariant Multilayer Perceptrons for Arbitrary Matrix Groupscited by 187
  4. 2024Don't Hallucinate, Abstain: Identifying LLM Knowledge Gaps via Multi-LLM Collaborationcited by 180
  5. 2024To Believe or Not to Believe Your LLMcited by 120
  6. 2024Do I Know This Entity? Knowledge Awareness and Hallucinations in Language Modelscited by 115
  7. 2024ClashEval: Quantifying the tug-of-war between an LLM's internal prior and external evidencecited by 113
  8. 2024FaithEval: Can Your Language Model Stay Faithful to Context, Even If "The Moon is Made of Marshmallows"cited by 86
  9. 2024Sufficient Context: A New Lens on Retrieval Augmented Generation Systemscited by 72
  10. 2024Rejection Improves Reliability: Training LLMs to Refuse Unknown Questions Using RL from Knowledge Feedbackcited by 66
  11. 2025Calibrating Verbal Uncertainty as a Linear Feature to Reduce Hallucinationscited by 48
  12. 2023Learn to Refuse: Making Large Language Models More Controllable and Reliable through Knowledge Scope Limitation and Refusal Mechanismcited by 33
  13. 2025Removal of Hallucination on Hallucination: Debate-Augmented RAGcited by 29
  14. 2024Rowen: Adaptive Retrieval-Augmented Generation for Hallucination Mitigation in LLMscited by 27
  15. 2024Characterizing LLM Abstention Behavior in Science QA with Context Perturbationscited by 26
  16. 2024Teaching LLMs to Abstain across Languages via Multilingual Feedbackcited by 14
  17. 2025Why Uncertainty Estimation Methods Fall Short in RAG: An Axiomatic Analysiscited by 14
  18. 2025Divide-Then-Align: Honest Alignment based on the Knowledge Boundary of RAGcited by 14
  19. 2024Controlling Risk of Retrieval-augmented Generation: A Counterfactual Prompting Frameworkcited by 12
  20. 2024Enabling LLM Knowledge Analysis via Extensive Materializationcited by 12
  21. 2024Mitigating Reversal Curse in Large Language Models via Semantic-aware Permutation Trainingcited by 11
  22. 2025MIRAGE: A Metric-Intensive Benchmark for Retrieval-Augmented Generation Evaluationcited by 11
  23. 2024Deciphering the Interplay of Parametric and Non-parametric Memory in Retrieval-augmented Language Modelscited by 10
  24. 2025Large Language-Geometry Model: When LLM meets Equivariancecited by 10
  25. 2024When to Speak, When to Abstain: Contrastive Decoding with Abstentioncited by 8
  26. 2025Unbiased Reasoning for Knowledge-Intensive Tasks in Large Language Models via Conditional Front-Door Adjustmentcited by 6
  27. 2025Hallucinate Less by Thinking More: Aspect-Based Causal Abstention for Large Language Modelscited by 5
  28. 2025Rethinking Addressing in Language Models via Contexualized Equivariant Positional Encodingcited by 4
  29. 2025On Entity Identification in Language Modelscited by 3
  30. 2025HALT-RAG: A Task-Adaptable Framework for Hallucination Detection with Calibrated NLI Ensembles and Abstentioncited by 3
  31. 2025HalluGuard: Evidence-Grounded Small Reasoning Models to Mitigate Hallucinations in Retrieval-Augmented Generationcited by 2
  32. 2026Debiasing Large Language Models via Adaptive Causal Prompting with Sketch-of-Thoughtcited by 2

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 1.

1 paper.

  1. 2026The Commit-Abstain Circuit: Why Language Models Hallucinate Instead of Abstainingcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Exploring Knowledge Conflicts for Faithful LLM Reasoning: Benchmark and Method
  2. 2026Knowing When Not to Answer: Abstention-Aware Scientific Reasoning
  3. 2026Knowledgeless Language Models: Suppressing Parametric Recall for Evidence-Grounded Language Modeling
  4. 2026DERELAB: Probing Defeasible Reasoning and Confirmation Bias in LLMs with a Generative Benchmark
  5. 2025Flying Pigs, FaR and Beyond: Evaluating LLM Reasoning in Counterfactual Worlds
  6. 2025Decoupling Knowledge and Reasoning in LLMs: An Exploration Using Cognitive Dual-System Theory
  7. 2026Double-Calibration: Towards Reliable LLMs via Calibrating Knowledge and Reasoning Confidence
  8. 2025Causal Sufficiency and Necessity Improves Chain-of-Thought Reasoning