{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreiedaxeuba2kv22izun2zbqitgighorcgc4ay2kunjo2pnwv63ljpu",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mpod2hmzuur2"
  },
  "path": "/t/research-from-functional-geometry-to-dynamic-grammar-new-limen-audits-v23-v24-across-7-architectures/177260#post_2",
  "publishedAt": "2026-07-02T13:16:03.000Z",
  "site": "https://discuss.huggingface.co",
  "tags": [
    "Markovian Circuit Tracing for Transformer State Dynamic",
    "PoLLMgraph: Unraveling Hallucinations in Large Language Models via State Transition Dynamics",
    "ICR Probe: Tracking Hidden State Dynamics for Reliable Hallucination Detection in LLMs",
    "code repository",
    "Transformers Represent Belief State Geometry in Their Residual Stream",
    "Eliciting Latent Predictions from Transformers with the Tuned Lens",
    "LogitLens4LLMs",
    "repo",
    "(click for more details)",
    "Transformers model output docs",
    "The hidden_states when I use model.generate",
    "Hidden states are different for model() and model.generate()",
    "Linear Probes Detect Task Format, Not Reasoning Mode in Language Model Hidden States",
    "Markovian Circuit Tracing",
    "PoLLMgraph",
    "ICR Probe",
    "Belief State Geometry",
    "Tuned Lens",
    "TransformerLens",
    "nnsight"
  ],
  "textContent": "I broadened the search terms and found a few leads that may be useful:\n\n* * *\n\nI would search less for **“symbolic dynamics of hidden states”** alone and more for terms like **activation-derived state abstraction** , **coarse state-transition structure** , **Markov/HMM-style hidden-state dynamics** , **internal state-transition dynamics during generation** , **cross-layer hidden-state dynamics** , and **belief-state geometry**.\n\nMy impression is that your V24 framing is adjacent to **state-transition abstractions over activation trajectories** , rather than ordinary probing alone. I would not treat the papers below as identical to LIMEN, but they seem like useful comparison points and search anchors.\n\n## 1. Nearby literature / comparison points\n\nArea | Why it seems relevant\n---|---\n**Markov/HMM-style activation-state abstraction** | Closest to the idea of recovering coarse internal states and transition motifs from activations.\n**Hallucination detection via state transitions** | Relevant to the B→D / grammar-violation idea.\n**Cross-layer hidden-state dynamics** | Relevant because the signal is not just a static hidden vector, but how representations evolve across layers.\n**Belief-state geometry** | A useful theoretical bridge from continuous representation geometry to coarse state dynamics.\n**Latent prediction trajectories** | Useful for checking whether a “D” state is actually decision-like / commitment-like.\n\n### Closest matches I found\n\n  1. **Markovian Circuit Tracing for Transformer State Dynamic**\nThis seems like the closest comparison point I found. It frames transformer computation as movement through internal states and tests whether activations contain **coarse state-transition structure**. The setup is synthetic HMM tasks where latent states, transition matrices, Bayesian belief vectors, Bayes-optimal predictions, and counterfactual targets are known.\n\nI would not say this is the same setting as your work, because it is a controlled HMM benchmark rather than natural-language ambiguity across open-weight LLMs. But as a reference framework for **activation-derived state abstractions + transition matrices + state forcing / centroid patching** , it seems highly relevant.\n\n  2. **PoLLMgraph: Unraveling Hallucinations in Large Language Models via State Transition Dynamics**\nThis seems especially relevant to your question about whether a B→D-like jump, or a violation of a B→A→D grammar, could be related to hallucination or reasoning failure. PoLLMgraph treats hallucination detection/forecasting as a problem of **internal state-transition dynamics during generation**.\n\n  3. **ICR Probe: Tracking Hidden State Dynamics for Reliable Hallucination Detection in LLMs** and its code repository\nThis is not a symbolic-transition-grammar paper, but it is relevant because it shifts from static hidden-state probes to the **cross-layer evolution of hidden states**. That seems adjacent to your layer-wise dynamic-regime framing.\n\n  4. **Transformers Represent Belief State Geometry in Their Residual Stream**\nThis seems like a useful theoretical bridge. It argues that, in HMM-like generative settings, transformer residual streams can linearly represent belief-state geometry. That does not directly prove anything about natural-language hallucination or B/A/D regimes, but it gives a vocabulary for connecting **functional geometry** to **coarse state dynamics**.\n\n  5. **Eliciting Latent Predictions from Transformers with the Tuned Lens**\nThis is not a state-transition grammar method, but it is useful for checking whether a state called “D” is actually decision-like. Tuned Lens decodes hidden states into vocabulary distributions layer by layer, so it can help compare state occupancy with prediction commitment.\n\n  6. **LogitLens4LLMs** / repo\nPotentially useful if you want a more implementation-oriented logit-lens style tool for modern HF models, including Qwen-family models.\n\n\nMore detailed relationship to your B/A/D framing (click for more details)\n\n## 2. Search terms that may help future readers\n\nA useful phrase might be:\n\n> **activation-derived coarse-state transition model**\n\nThat wording avoids over-committing to “symbolic grammar,” while still capturing the useful part of the idea: you are turning continuous activation trajectories into discrete-ish regimes and then studying transitions among those regimes.\n\nOther search terms I would try:\n\n  * activation-derived state abstraction\n  * coarse state-transition structure\n  * Markovian hidden-state dynamics\n  * HMM-style activation dynamics\n  * internal state-transition dynamics during generation\n  * cross-layer hidden-state dynamics\n  * belief-state geometry in residual streams\n  * latent prediction trajectories\n  * layer-wise prediction refinement\n  * hidden-state hallucination detection\n  * activation trajectory clustering\n  * state-transition interpretability\n  * fixed-layer time-transition analysis\n  * hidden-state dynamics hallucination detection\n\n\n\n## 3. One thing I would separate: clustering vs naming\n\nIf B/A/D are meant as functional names rather than just cluster labels, I would separate two steps:\n\n  1. **State construction**\nHow are the states obtained?\nClustering? Thresholding? A linear probe? A hand-built rule? A model-specific classifier?\n\n  2. **State naming**\nWhy is a state called exploration, stabilization, decision, etc.?\nWhat observable behavior supports that name?\n\n\n\n\nFor example, if **D = decision-like** , I would expect some relation to prediction commitment, such as:\n\n  * lower next-token entropy,\n  * larger top-1 / top-2 margin,\n  * higher agreement with final-layer predictions,\n  * earlier/later commitment timing,\n  * possibly a Tuned Lens / Logit Lens trajectory becoming more stable.\n\n\n\nBut if D can be predicted almost entirely from entropy or top-token margin, then “D” may still be useful, but the interpretation shifts toward **confidence proxy** rather than an independent functional state.\n\nPossible naming checks for B/A/D (click for more details)\n\n## 4. A small routing diagnostic I tried\n\nI also tried a small Colab-T4-style routing check. I would not treat this as validation or refutation of the dynamic-grammar idea. The point was only to see what simple state abstractions tend to pick up.\n\nThe setup was intentionally lightweight:\n\n  * small public models,\n  * manual greedy decoding,\n  * hidden states collected with `output_hidden_states=True`,\n  * raw vs layer-centered vs layer-z-scored hidden states,\n  * PCA/KMeans-style state abstraction,\n  * checks against depth-only, prompt-family, entropy/margin, generation-step, shuffled-layer, and fixed-layer time-transition baselines.\n\n\n\nThe most useful observation was:\n\n> Raw layer-wise transition motifs can look clean while being strongly aligned with layer depth.\n\nLayer centering or z-scoring can reduce that signal in some models, but it can also expose other proxy alignments such as prompt family, entropy/margin bins, or generation step. So I would frame these checks as **routing diagnostics** , not as an objection to the dynamic-grammar idea.\n\nWhat the routing diagnostic suggested (click for more details)\n\n## 5. Suggested reporting table\n\nIf you continue this line of work, I think a compact table like this would make the claim easier to compare across models and future replications.\n\nView | What it tests | Useful controls\n---|---|---\nraw layer-wise states | Original observed motif | shuffled-layer, random Gaussian\nlayer-centered states | Non-depth residual structure | depth-only, prompt-family\nlayer-z-scored states | Mean/scale controlled structure | entropy/margin, step\nfixed-layer time transitions | Generation-time dynamics without layer-depth transitions | shuffled-time, prompt-family\nD-state checks | Whether “D” is decision-like | entropy, margin, tuned/logit lens\nB→D violation checks | Whether transition skips predict failure | hallucination / reasoning-error cases\nheld-out prompt families | Generality beyond prompt template | prompt-family split\nheld-out model families | Generality beyond one architecture | model-family split\n\nA minimal version could be:\n\n\n    raw / layer-centered / layer-z-scored\n    ×\n    depth / prompt-family / entropy / margin / step / shuffled-order / fixed-layer time-transition\n\n\nThat would make the B/A/D interpretation much easier to evaluate without requiring a huge model run.\n\n## 6. Reproducibility note: hidden-state extraction contract\n\nNot as an objection to your result, but for this kind of geometry/transition analysis, I think the exact hidden-state extraction contract matters a lot.\n\nThe HF docs note that `outputs.hidden_states[-1]` may not always match `outputs.last_hidden_state`, because some models apply normalization or other processing to the returned last hidden state: Transformers model output docs.\n\nGeneration-time hidden states also need care. There is an HF Forum thread on retrieving hidden states from `model.generate()` where the answer separates generated-token states from input-sequence states and mentions `return_dict_in_generate=True` / `output_hidden_states=True`: The hidden_states when I use model.generate.\n\nThere is also a Transformers issue where `model()` and `model.generate()` hidden states did not match as expected in a Qwen2.5 example: Hidden states are different for model() and model.generate(). I would not read that issue as applying directly to your setup, but it is a good reminder that extraction details should be explicit.\n\nExtraction details I would report (click for more details)\n\n## 7. Probe / format confound caution\n\nOne related caution is that high hidden-state separability does not automatically imply a functional internal variable. A recent example is Linear Probes Detect Task Format, Not Reasoning Mode in Language Model Hidden States, where high probe accuracy on Qwen3-14B was explained by format confounds such as source identity, option count, and response length.\n\nI would not treat that as a refutation of your result. I would treat it as a useful reminder to include format/depth/confidence controls before assigning strong functional interpretations to state labels.\n\nIn this context, the main question becomes:\n\n> Are B/A/D labels still informative after controlling for layer depth, prompt family, response format, token position, entropy, top-token margin, and generation step?\n\nIf yes, the dynamic-grammar interpretation becomes easier to compare.\nIf no, the result may still be useful, but the interpretation shifts toward depth/confidence/prompt-format dynamics.\n\n## 8. Possible next tracks\n\nI would split the next steps into three tracks.\n\n### Track A — terminology / literature mapping\n\nUse these as comparison anchors:\n\n  * Markovian Circuit Tracing for activation-state abstraction and transition matrices.\n  * PoLLMgraph for hallucination as internal state-transition dynamics.\n  * ICR Probe for cross-layer hidden-state evolution.\n  * Belief State Geometry for the continuous-geometry side.\n  * Tuned Lens / LogitLens4LLMs for prediction-trajectory checks.\n\n\n\n### Track B — observational controls\n\nBefore moving to large models or causal claims, I would add a small diagnostic table:\n\n  * raw vs layer-centered vs layer-z-scored states,\n  * depth-only baseline,\n  * entropy/margin-only baseline,\n  * prompt-family baseline,\n  * generation-step baseline,\n  * shuffled-layer transition matrix,\n  * random Gaussian baseline,\n  * fixed-layer time-transition matrix.\n\n\n\nThis is relatively cheap and makes the interpretation much clearer.\n\n### Track C — later predictive / causal validation\n\nOnce the observational state definition is stable:\n\n  * test whether B→D-like jumps predict factual hallucination or reasoning errors,\n  * compare correct vs incorrect generations,\n  * try held-out prompt families,\n  * try model-family-held-out mapping,\n  * eventually test centroid patching / activation patching.\n\n\n\nFor tooling, TransformerLens is useful for caching and editing activations in supported models, and nnsight has activation-patching style workflows. I would treat that as a later stage, after the observational labels are stable.\n\n## 9. How I would frame the claim\n\nI would avoid framing the current result as either:\n\n  * “a universal transformer grammar,” or\n  * “just a probing artifact.”\n\n\n\nA safer and more useful middle formulation might be:\n\n> Under this extraction and labeling setup, there appear to be recurrent activation-derived transition motifs. The next step is to report which motifs survive depth, confidence, prompt-family, step, shuffled-order, and fixed-layer controls.\n\nThat keeps the core idea alive while making it easier for other people to reproduce, compare, and extend.\n\nIn short: the B/A/D grammar idea seems worth pursuing, but I would make the next iteration less about proving the names and more about making the state abstraction auditable.",
  "title": "[Research] From Functional Geometry to Dynamic Grammar: New LIMEN Audits (V23–V24) Across 7 Architectures"
}