mlsanityData to 7 Oct 2026

Overcoming State Inertia: Minimally Invasive Temporal Alignment for Evolving Contexts

arXiv 2512.03704 ↗First posted 3 Dec 2025cs.CLcited by 1

On the map: LLM alignment / RLHF / preference optimisation → LLM preference optimization methods

Abstract (arXiv)

Long-context dialogue systems suffer from state inertia, where models over-attend to history and fail to adapt to evolving intents. We demonstrate that standard alignment methods like DPO and even recent long-context optimization techniques struggle to resolve this without incurring a severe contextual alignment tax--a substantial perplexity surge caused by disrupting pre-trained priors. To address this, we propose DZ-TiDPO, a minimally invasive framework that synergizes conflict-aware optimization (during training) with a structural temporal attention bias. This design effectively decouples state updating from general linguistic modeling. Experiments on Multi-Session Chat and our new Inertia Challenge (IC-Bench) show DZ-TiDPO preserves structural coherence while resolving inter-turn conflicts. Crucially, our framework supports dual inference strategies: a negligible-latency static mode for general robustness and a precision-focused dynamic mode for micro-semantic conflicts. Furthermore, our scaling analysis reveals a capacity-stability trade-off, confirming that highly capable mid-sized models (7B) can efficiently internalize temporal alignment. Code and data are available at: https://github.com/lyj20071013/DZ-TiDPO.

Key facts. Cited by 1 paper in the map (data through 7 Oct 2026). Placed in LLM preference optimization methods, one of 10 directions in LLM alignment / RLHF / preference optimisation. Builds on 52 of its 87 references that are in the map. First posted on arXiv 3 Dec 2025.

Builds on

52 of 87 references.

Closest ancestors · 8

Same direction, newest first

  1. 2025SoLoPO: Unlocking Long-Context Capabilities in LLMs via Short-to-Long Preference Optimization16 May
  2. 2025LongPO: Long Context Self-Evolution of Large Language Models through Short-to-Long Preference Optimization19 Feb
  3. 2024SimPO: Simple Preference Optimization with a Reference-Free Reward23 May
  4. 2024ORPO: Monolithic Preference Optimization without Reference Model12 Mar
  5. 2023A General Theoretical Paradigm to Understand Learning from Human Preferences18 Oct
  6. 2023Mitigating the Alignment Tax of RLHF12 Sep
  7. 2022Training language models to follow instructions with human feedback4 Mar
  8. 2021A General Language Assistant as a Laboratory for Alignment1 Dec

Other references · 44

  1. 2020Retrieval-Augmented Generation for Knowledge-Intensive NLP Taskscited by 9,160
  2. 2019Sentence-BERT: Sentence Embeddings using Siamese BERT-Networkscited by 9,049
  3. 2023Judging LLM-as-a-Judge with MT-Bench and Chatbot Arenacited by 8,393
  4. 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 7,916
  5. 2020Measuring Massive Multitask Language Understandingcited by 7,768
  6. 2019BERTScore: Evaluating Text Generation with BERTcited by 5,462
  7. 2023Mamba: Linear-Time Sequence Modeling with Selective State Spacescited by 4,616
  8. 2021RoFormer: Enhanced Transformer with Rotary Position Embeddingcited by 4,581
  9. 2023Universal and Transferable Adversarial Attacks on Aligned Language Modelscited by 2,914
  10. 2020Learning to summarize from human feedbackcited by 2,562
  11. 2023Efficient Streaming Language Models with Attention Sinkscited by 2,115
  12. 2024Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phonecited by 1,883
32 more
  1. 2018A Call for Clarity in Reporting BLEU Scorescited by 1,790
  2. 2021WebGPT: Browser-assisted question-answering with human feedbackcited by 1,722
  3. 2020MiniLM: Deep Self-Attention Distillation for Task-Agnostic Compression of Pre-Trained Transformerscited by 1,414
  4. 2024From Local to Global: A Graph RAG Approach to Query-Focused Summarizationcited by 1,411
  5. 2023MemGPT: Towards LLMs as Operating Systemscited by 1,295
  6. 2024RULER: What's the Real Context Size of Your Long-Context Language Models?cited by 1,270
  7. 2023Enhancing Chat Language Models by Scaling High-quality Instructional Conversationscited by 825
  8. 2023RLAIF vs. RLHF: Scaling Reinforcement Learning from Human Feedback with AI Feedbackcited by 615
  9. 2024Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Modelscited by 552
  10. 2024RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrievalcited by 519
  11. 2024Quest: Query-Aware Sparsity for Efficient Long-Context LLM Inferencecited by 487
  12. 2024From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judgecited by 480
  13. 2023Ring Attention with Blockwise Transformers for Near-Infinite Contextcited by 448
  14. 2024The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructionscited by 441
  15. 2024PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funnelingcited by 420
  16. 2023The Reversal Curse: LLMs trained on "A is B" fail to learn "B is A"cited by 399
  17. 2024Titans: Learning to Memorize at Test Timecited by 303
  18. 2024Quiet-STaR: Language Models Can Teach Themselves to Think Before Speakingcited by 292
  19. 2024DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Headscited by 283
  20. 2021Beyond Goldfish Memory: Long-Term Open-Domain Conversationcited by 266
  21. 2013Non-stationary Stochastic Optimizationcited by 173
  22. 2025Safety Tax: Safety Alignment Makes Your Large Reasoning Models Less Reasonablecited by 120
  23. 2024Aligning LLMs with Individual Preferences via Interactioncited by 63
  24. 2024Language Models Resist Alignment: Evidence From Data Compressioncited by 50
  25. 2025RocketKV: Accelerating Long-Context LLM Inference via Two-Stage KV Cache Compressioncited by 41
  26. 2025(How) Do Language Models Track State?cited by 31
  27. 2025Persistent Instability in LLM's Personality Measurements: Effects of Scale, Reasoning, and Conversation Historycited by 25
  28. 2025Multi-Agent Debate for LLM Judges with Adaptive Stability Detectioncited by 20
  29. 2024Contextual Document Embeddingscited by 19
  30. 2024System 2 thinking in OpenAI's o1-preview model: Near-perfect performance on a mathematics examcited by 10
  31. 2025Are LLM Belief Updates Consistent with Bayes' Theorem?cited by 8
  32. 2025Semantic Adapter for Universal Text Embeddings: Diagnosing and Mitigating Negation Blindness to Enhance Universalitycited by 1

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 1.

1 paper.

  1. 2026Shortcut Before Circuit: Document Statistics Time In-Context Conflict Resolutioncited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026ACR: Adaptive Context Refactoring via Context Refactoring Operators for Multi-Turn Dialogue
  2. 2026A Context Alignment Pre-processor for Enhancing the Coherence of Human-LLM Dialog
  3. 2025D-SMART: Enhancing LLM Dialogue Consistency via Dynamic Structured Memory And Reasoning Tree
  4. 2026Context-Agent: Dynamic Discourse Trees for Non-Linear Dialogue
  5. 2025Aligning Spoken Dialogue Models from User Interactions
  6. 2023OpenAssistant Conversations -- Democratizing Large Language Model Alignment
  7. 2023In-Context Alignment: Chat with Vanilla Language Models Before Fine-Tuning
  8. 2026Beyond Surface Alignment: Grounding the Dynamics of Situational Understanding and Generative Control in LLMs