Causal Interventions in Language Models

Across multiple language pairings (English $\to$ {Spanish, French, German, Mandarin, Japanese, Urdu, Cantonese}), we find reasoning errors in translation. To quantify how often these reasoning errors occur, we leverage an automated annotation protocol for reasoning evaluation wherein the goal is to detect if a reasoning step is any of three error categories: (1) source sentence-misaligned, (2) model hypothesis-misaligned, or (3) reasoning trace-misaligned. We probe the reasoning model with perturbed traces correcting for these identified reasoning errors using an array of weak-to-strong interventions: hedging, removal, re-reasoning after removal, hindsight, and oracle interventions. Experimenting with interventions on the reasoning traces suggests that small corrections to the reasoning have little impact on translation quality, but stronger interventions yield the highest resolution rates, despite translation quality gains being mixed. We find ultimately that reasoning errors in MT can be identified with high precision in Urdu but lower precision in Spanish, but that removing these reasoning errors does not resolve the initial errors significantly, suggesting limited reasoning faithfulness for machine translation.

Understanding how large language models encode task identity from few-shot demonstrations is a central open problem in mechanistic interpretability. Prior work uses linear probing to localize task representations, reporting high classification accuracy at specific layers. We reveal a striking dissociation: probing accuracy completely fails to predict causal importance. Single-position activation intervention achieves 0% task transfer across all 28 layers of Llama-3.2-3B-despite 100% probing accuracy at those same positions. This null result is itself a key finding, demonstrating that task encoding is fundamentally distributed. Multi-position intervention-replacing activations at all demonstration output tokens simultaneously-achieves up to 96% transfer (N=50, 95% CI: [87%, 99%]) at layer 8, pinpointing for the first time the causal locus of ICL task identity. We establish the generality of these findings across four models spanning three architecture families (LLaMA, Qwen, Gemma), discovering a universal intervention window at ~30% network depth. Causal tracing uncovers an asymmetric architecture: the query position is strictly necessary (53-100% disruption) while no individual demonstration position is necessary (0% disruption)-resolving a key ambiguity in prior accounts. Crucially, transfer depends on internal representation compatibility, not surface similarity (r=-0.05 vs r=0.31), ruling out trivial explanations. These results establish the distributed template hypothesis: ICL task identity is encoded as output format templates distributed across demonstration tokens, fundamentally reshaping our understanding of how in-context learning operates.

Large language models store not only isolated facts but also rules that support reasoning across symbolic expressions, natural language explanations, and concrete instances. Yet most model editing methods are built for fact-level knowledge, assuming that a target edit can be achieved through a localized intervention. This assumption does not hold for rule-level knowledge, where a single rule must remain consistent across multiple interdependent forms. We investigate this problem through a mechanistic study of rule-level knowledge editing. To support this study, we extend the RuleEdit benchmark from 80 to 200 manually verified rules spanning mathematics and physics. Fine-grained causal tracing reveals a form-specific organization of rule knowledge in transformer layers: formulas and descriptions are concentrated in earlier layers, while instances are more associated with middle layers. These results suggest that rule knowledge is not uniformly localized, and therefore cannot be reliably edited by a single-layer or contiguous-block intervention. Based on this insight, we propose Distributed Multi-Layer Editing (DMLE), which applies a shared early-layer update to formulas and descriptions and a separate middle-layer update to instances. While remaining competitive on standard editing metrics, DMLE achieves substantially stronger rule-level editing performance. On average, it improves instance portability and rule understanding by 13.91 and 50.19 percentage points, respectively, over the strongest baseline across GPT-J-6B, Qwen2.5-7B, Qwen2-7B, and LLaMA-3-8B. The code is available at https://github.com/Pepper66/DMLE.

While Large Language Models (LLMs) have achieved remarkable performance, they remain vulnerable to jailbreak attacks that circumvent safety constraints. Existing strategies, ranging from heuristic prompt engineering to computationally intensive optimization, often face significant trade-offs between effectiveness and efficiency. In this work, we propose Contextual Representation Ablation (CRA), a novel inference-time intervention framework designed to dynamically silence model guardrails. Predicated on the geometric insight that refusal behaviors are mediated by specific low-rank subspaces within the model's hidden states, CRA identifies and suppresses these refusal-inducing activation patterns during decoding without requiring expensive parameter updates or training. Empirical evaluation across multiple safety-aligned open-source LLMs demonstrates that CRA significantly outperforms baselines. These results expose the intrinsic fragility of current alignment mechanisms, revealing that safety constraints can be surgically ablated from internal representations, and underscore the urgent need for more robust defenses that secure the model's latent space.

Large language models often fail on tasks they seem to already understand. In our experiments, this appears to be less about missing knowledge and more about certain internal circuits not being strongly activated during inference. We explore Selective Neuron Amplification, which increases the influence of task relevant neurons without changing the model's parameters. The method works at inference time and does not permanently alter the model. SNA helps mainly when the model is uncertain, while having low effect when the model is already confident. This suggests that some model failures are due to weak activation rather than lack of capability.

Pragmatic inference is inherently graded. Different lexical items give rise to pragmatic enrichment to different degrees. Scalar implicature exemplifies this property through scalar diversity, where implicature strength varies across scalar items. However, evaluations of pragmatic inference in large language models (LLMs) often rely on prompt-based manipulations. Beyond prompt-level effects, this study introduces Continuous Interpretive Steering (CIS), a method that probes graded pragmatic interpretation by treating activation-level steering strength as a continuous experimental variable. To support this analysis, this study introduces a new dataset, GraSD, which encodes graded scalar diversity. Experiments on four LLMs show that uniform activation steering increases pragmatic interpretations globally but collapses item-level variation, whereas graded activation steering yields differentiated interpretive shifts aligned with scalar diversity grades. It indicates that graded sensitivity is encoded in the representation space and can be systematically recovered through controlled intervention. Together, CIS and GraSD provide a principled framework for evaluating graded pragmatic sensitivity in LLMs.

While large language models (LLMs) improve performance by explicit reasoning, their responses are often overconfident, even though they include linguistic expressions demonstrating uncertainty. In this work, we identify what linguistic expressions are related to confidence by applying the regression method. Specifically, we predict confidence of those linguistic expressions in the reasoning parts of LLMs as the dependent variables and analyze the relationship between a specific $n$-gram and confidence. Across multiple models and QA benchmarks, we show that LLMs remain overconfident when reasoning is involved and attribute this behavior to specific linguistic information. Interestingly, several of the extracted expressions coincide with cue phrases intentionally inserted on test-time scaling to improve reasoning performance. Through our test on causality and verification that the extracted linguistic information truly affects confidence, we reveal that confidence calibration is possible by simply suppressing those overconfident expressions without drops in performance.

As language models shift from single-shot answer generation toward multi-step reasoning that retrieves and consumes evidence mid-inference, evaluating the role of individual retrieved items becomes more important. Existing RAG evaluation typically targets final-answer quality, citation faithfulness, or answer-level attribution, but none of these directly targets the intervention-based, per-evidence-item utility view we study here. We introduce CUE-R, a lightweight intervention-based framework for measuring per-evidence-item operational utility in single-shot RAG using shallow observable retrieval-use traces. CUE-R perturbs individual evidence items via REMOVE, REPLACE, and DUPLICATE operators, then measures changes along three utility axes (correctness, proxy-based grounding faithfulness, and confidence error) plus a trace-divergence signal. We also outline an operational evidence-role taxonomy for interpreting intervention outcomes. Experiments on HotpotQA and 2WikiMultihopQA with Qwen-3 8B and GPT-5.2 reveal a consistent pattern: REMOVE and REPLACE substantially harm correctness and grounding while producing large trace shifts, whereas DUPLICATE is often answer-redundant yet not fully behaviorally neutral. A zero-retrieval control confirms that these effects arise from degradation of meaningful retrieval. A two-support ablation further shows that multi-hop evidence items can interact non-additively: removing both supports harms performance far more than either single removal. Our results suggest that answer-only evaluation misses important evidence effects and that intervention-based utility analysis is a practical complement for RAG evaluation.

Models can represent policy-relevant content yet respond differently. We study the policy-routing circuit that turns detection into refusal. Using political censorship and harmful-request refusal as test cases, four-method circuit analyses in Qwen3-8B, Phi-4-mini, and Gemma-2-2B reveal sparse gate-amplifier circuits at architecture-specific locations. An intermediate-layer gate initiates routing, while later, distributed amplifier heads strengthen a refusal-directed signal. The gate is hard to find with direct logit attribution (DLA): in Qwen3-8B, it contributes less than 1% of output DLA. We therefore combine DLA screening, directional intervention, activation interchange, and gate-to-amplifier knockout. Activation interchange identifies upstream trigger candidates, and knockout suppresses a majority of tested later-head contributions in all three models. The gate is an upstream trigger, not a self-contained refusal mechanism; later amplifiers strengthen the signal toward the output. Candidate-head screens in nine additional models from six labs, spanning 2B-72B parameters, nominate similarly localized routing candidates without establishing head-level homology. Policy routing is a sparse, causally structured interface between content detection and aligned behavior that output attribution alone can miss.

This paper primarily focuses on the hallucinations caused due to AI language models(LLMs).LLMs have shown extraordinary Language understanding and generation capabilities .Still it has major a disadvantage hallucinations which give outputs which are factually incorrect ,misleading or unsupported by input data . These hallucinations cause serious problems in scenarios like medical diagnosis or legal reasoning.Through this work,we propose causal graph attention network (GCAN) framework that reduces hallucinations through interpretation of internal attention flow within a transformer architecture with the help of constructing token level graphs that combine self attention weights and gradient based influence scores.our method quantifies each tokens factual dependency using a new metric called the Causal Contribution Score (CCS). We further introduce a fact-anchored graph reweighting layer that dynamically reduces the influence of hallucination prone nodes during generation. Experiments on standard benchmarks such as TruthfulQA and HotpotQA show a 27.8 percent reduction in hallucination rate and 16.4 percent improvement in factual accuracy over baseline retrieval-augmented generation (RAG) models. This work contributes to the interpretability,robustness, and factual reliability of future LLM architectures.

Large language models demonstrate strong performance on mathematical reasoning benchmarks, yet remain surprisingly fragile to meaning-preserving surface perturbations. We systematically evaluate three open-weight LLMs, Mistral-7B, Llama-3-8B, and Qwen2.5-7B, on 677 GSM8K problems paired with semantically equivalent variants generated through name substitution and number format paraphrasing. All three models exhibit substantial answer-flip rates (28.8%-45.1%), with number paraphrasing consistently more disruptive than name swaps. To trace the mechanistic basis of these failures, we introduce the Mechanistic Perturbation Diagnostics (MPD) framework, combining logit lens analysis, activation patching, component ablation, and the Cascading Amplification Index (CAI) into a unified diagnostic pipeline. CAI, a novel metric quantifying layer-wise divergence amplification, outperforms first divergence layer as a failure predictor for two of three architectures (AUC up to 0.679). Logit lens reveals that flipped samples diverge from correct predictions at significantly earlier layers than stable samples. Activation patching reveals a stark architectural divide in failure localizability: Llama-3 failures are recoverable by patching at specific layers (43/60 samples), while Mistral and Qwen failures are broadly distributed (3/60 and 0/60). Based on these diagnostic signals, we propose a mechanistic failure taxonomy (localized, distributed, and entangled) and validate it through targeted repair experiments: steering vectors and layer fine-tuning recover 12.2% of localized failures (Llama-3) but only 7.2% of entangled (Qwen) and 5.2% of distributed (Mistral) failures.

Large language models are often not just wrong, but confidently wrong: when they produce factually incorrect answers, they tend to verbalize overly high confidence rather than signal uncertainty. Such verbalized overconfidence can mislead users and weaken confidence scores as a reliable uncertainty signal, yet its internal mechanisms remain poorly understood. We present a circuit-level mechanistic analysis of this inflated verbalized confidence in LLMs, organized around three axes: capturing verbalized confidence as a differentiable internal signal, identifying the circuits that causally inflate it, and leveraging these insights for targeted inference-time recalibration. Across two instruction-tuned LLMs on three datasets, we find that a compact set of MLP blocks and attention heads, concentrated in middle-to-late layers, consistently writes the confidence-inflation signal at the final token position. We further show that targeted inference-time interventions on these circuits substantially improve calibration. Together, our results suggest that verbalized overconfidence in LLMs is driven by identifiable internal circuits and can be mitigated through targeted intervention.

How do language models retrieve entity-specific facts from their parameters? We investigate this question by searching for sparse, entity-selective MLP neurons - which we call entity cells, by analogy to the "grandmother cell" hypothesis in neuroscience - and testing whether they play a causal role in factual recall. We localize candidate entity cells by ranking MLP neurons for activation consistency across varied prompts about the same entity, applying this procedure across seven models on a curated subset of PopQA. In all models, localized neurons cluster predominantly in early layers, an empirical pattern not imposed by the architecture. Using Qwen2.5-7B base as a model organism, we find the clearest causal evidence: suppressing a localized cell selectively erases recall for its matched entity while leaving others intact, and activating a single cell is sufficient to recover correct knowledge for most entities - even when the entity is absent from the context. The same cells are recovered under aliases, acronyms, misspellings, and multilingual surface forms, and remain stable through instruction tuning, suggesting they encode canonical entity identity rather than surface token patterns. Causal signals vary across model families, pointing to architectural differences in how entity knowledge is organized. These findings offer concrete, interpretable access points for understanding, controlling, and correcting factual knowledge in language models, and draw a surprising empirical parallel to longstanding questions in neuroscience about sparse coding of concepts.

Routing is widely used to scale large language models, from Mixture-of-Experts gating to multi-model/tool selection. A common belief is that routing to a task ``expert'' activates sparser internal computation and thus yields more certain and stable outputs (the Sparsity--Certainty Hypothesis). We test this belief by injecting routing-style meta prompts as a textual proxy for routing signals in front of frozen instruction-tuned LLMs. We quantify (C1) internal density via activation sparsity, (C2) domain-keyword attention, and (C3) output stability via predictive entropy and semantic variation. On a RouterEval subset with three instruction-tuned models (Qwen3-8B, Llama-3.1-8B-Instruct, and Mistral-7B-Instruct-v0.2), meta prompts consistently densify early/middle-layer representations rather than increasing sparsity; natural-language expert instructions are often stronger than structured tags. Attention responses are heterogeneous: Qwen/Llama reduce keyword attention, while Mistral reinforces it. Finally, the densification--stability link is weak and appears only in Qwen, with near-zero correlations in Llama and Mistral. We present RIDE as a diagnostic probe for calibrating routing design and uncertainty estimation.

System prompt instructions that cooperate in English compete in Spanish, with the same semantic content, but opposite interaction topology. We present instruction-level ablation experiments across four languages and four models showing that this topology inversion is mediated by social register: the imperative mood carries different obligatory force across speech communities, and models trained on multilingual data have learned these conventions. Declarative rewriting of a single instruction block reduces cross-linguistic variance by 81% (p = 0.029, permutation test). Rewriting three of eleven imperative blocks shifts Spanish instruction topology from competitive to cooperative, with spillover effects on unrewritten blocks. These findings suggest that models process instructions as social acts, not technical specifications: "NEVER do X" is an exercise of authority whose force is language-dependent, while "X: disabled" is a factual description that transfers across languages. If register mediates instruction-following at inference time, it plausibly does so during training. We state this as a testable prediction: constitutional AI principles authored in imperative mood may create language-dependent alignment. Corpus: 22 hand-authored probes against a production system prompt decomposed into 56 blocks.

Language models increasingly show their work by writing step-by-step reasoning before answering. But are these steps genuinely used, or is the answer rigid - fixed before reasoning begins? We introduce the Step-Level Reasoning Capacity (SLRC) metric and prove it is a consistent causal estimator (Theorem 1). We propose LC-CoSR, a training method with Lyapunov stability guarantees that directly reduces rigidity. Evaluating 16 frontier models (o4-mini, GPT-5.4, Claude Opus, Grok-4, DeepSeek-R1, Gemini 2.5 Pro, and others) across six domains at N=133-500, we find reasoning falls into three modes. OpenAI's o4-mini shows 74-88% step necessity on five of six tasks (73.8-88.3%) - the highest SLRC in our study. The critical differentiator is RL-based reasoning training, not thinking tokens: Grok-4's reasoning mode shows lower faithfulness than its non-reasoning mode (1.4% vs 7.2% necessity). We discover a faithfulness paradox - high-SLRC models are more susceptible to sycophancy - and propose the Reasoning Integrity Score (RIS = SLRC x (1-Sycophancy)), which significantly predicts error detection (rho=0.66, p=0.026). LC-CoSR achieves 2.6x less negative reward than FARL and CSR baselines without external model dependencies.

Reinforcement learning with verifiable rewards (RLVR) has significantly improved reasoning in large language models (LLMs), yet the token-level mechanisms underlying these improvements remain unclear. We present a systematic empirical study of RLVR's distributional effects organized around three main analyses: (1) token-level characterization of distributional shifts between base and RL models, (2) the impact of token-level distributional shifts on sequence-level reasoning performance through cross-sampling interventions, and (3) fine-grained mechanics of these shifts at the token level. We find that RL fine-tuning induces highly sparse and targeted changes, with only a small fraction of token distributions exhibiting meaningful divergence between the base and RL policies. We further characterize the structure and evolution of these shifts through analyses of token entropy, positional concentration, and reallocation of probability mass. To assess the functional importance of these sparse changes, we conduct cross-sampling experiments that selectively swap token choices between the base and RL models with varying intervention budgets. We show that inserting only a small fraction of RL-sampled tokens into base generations progressively recovers RL performance gains, while injecting a similarly small number of base token choices into otherwise RL-generated sequences collapses performance to base levels, isolating a small set of token-level decisions directly responsible for RLVR's performance gains. Finally, we explore divergence-weighted variants of the advantage signal as a diagnostic intervention, finding that they can yield improvements over baselines. Together, our results shed light on the distributional changes induced by RLVR and provide a fine-grained, token-level lens for understanding RLVR fine-tuning as a targeted refinement process.

Inasmuch as the removal of refusal behavior from instruction-tuned language models by directional abliteration requires the extraction of refusal-mediating directions from the residual stream activation space, and inasmuch as the construction of the contrast baseline against which harmful prompt activations are compared has been treated in the existing literature as an implementation detail rather than a methodological concern, the present work investigates whether a topically matched contrast baseline yields superior refusal directions. The investigation is carried out on the Qwen~3.5 2B model using per-category matched prompt pairs, per-class Self-Organizing Map extraction, and Singular Value Decomposition orthogonalization. It was found that topic-matched contrast produces no functional refusal directions at any tested weight level on any tested layer, while unmatched contrast on the same model, same extraction code, and same evaluation protocol achieves complete refusal elimination on six layers. The geometric analysis of the failure establishes that topic-matched subtraction cancels the dominant activation component shared between harmful and harmless prompts of the same subject, reducing the extracted direction magnitude below the threshold at which weight-matrix projection perturbs the residual stream. The implications for the design of contrast baselines in abliteration research are discussed.

Large language models (LLMs) are increasingly used as knowledge bases, but keeping them up to date requires targeted knowledge editing (KE). However, it remains unclear how edits are implemented inside the model once applied. In this work, we take a mechanistic view of KE using neuron-level knowledge attribution (NLKA). Unlike prior work that focuses on pre-edit causal tracing and localization, we use post-edit attribution -- contrasting successful and failed edits -- to isolate the computations that shift when an edit succeeds. Across representative KE methods, we find a consistent pattern: mid-to-late attention predominantly promotes the new target, while attention and FFN modules cooperate to suppress the original fact. Motivated by these findings, we propose MEGA, a MEchanism-Guided Activation steering method that performs attention-residual interventions in attribution-aligned regions without modifying model weights. On CounterFact and Popular, MEGA achieves strong editing performance across KE metrics on GPT2-XL and LLaMA2-7B. Overall, our results elevate post-edit attribution from analysis to engineering signal: by pinpointing where and how edits take hold, it powers MEGA to deliver reliable, architecture-agnostic knowledge edits.

Direct Preference Optimization (DPO) guides large language models (LLMs) to generate recommendations aligned with user historical behavior distributions by minimizing preference alignment loss. However, our systematic empirical research and theoretical analysis reveal that DPO tends to amplify spurious correlations caused by environmental confounders during the alignment process, significantly undermining the generalization capability of LLM-based generative recommendation methods in out of distribution (OOD) scenarios. To mitigate this issue, we propose CausalDPO, an extension of DPO that incorporates a causal invariance learning mechanism. This method introduces a backdoor adjustment strategy during the preference alignment phase to eliminate interference from environmental confounders, explicitly models the latent environmental distribution using a soft clustering approach, and enhances robust consistency across diverse environments through invariance constraints. Theoretical analysis demonstrates that CausalDPO can effectively capture users stable preference structures across multiple environments, thereby improving the OOD generalization performance of LLM-based recommendation models. We conduct extensive experiments under four representative distribution shift settings to validate the effectiveness of CausalDPO, achieving an average performance improvement of 17.17% across four evaluation metrics.

Can we trust the reasoning traces that large reasoning models (LRMs) produce? We investigate whether these traces faithfully reflect what drives model outputs, and whether models will honestly report their influence. We introduce Thought Injection, a method that injects synthetic reasoning snippets into a model's <think> trace, then measures whether the model follows the injected reasoning and acknowledges doing so. Across 45,000 samples from three LRMs, we find that injected hints reliably alter outputs, confirming that reasoning traces causally shape model behavior. However, when asked to explain their changed answers, models overwhelmingly refuse to disclose the influence: overall non-disclosure exceeds 90% for extreme hints across 30,000 follow-up samples. Instead of acknowledging the injected reasoning, models fabricate aligned-appearing but unrelated explanations. Activation analysis reveals that sycophancy- and deception-related directions are strongly activated during these fabrications, suggesting systematic patterns rather than incidental failures. Our findings reveal a gap between the reasoning LRMs follow and the reasoning they report, raising concern that aligned-appearing explanations may not be equivalent to genuine alignment.

Precise behavioral control of large language models (LLMs) is critical for complex applications. However, existing methods often incur high training costs, lack natural language controllability, or compromise semantic coherence. To bridge this gap, we propose WASD (unWeaving Actionable Sufficient Directives), a novel framework that explains model behavior by identifying sufficient neural conditions for token generation. Our method represents candidate conditions as neuron-activation predicates and iteratively searches for a minimal set that guarantees the current output under input perturbations. Experiments on SST-2 and CounterFact with the Gemma-2-2B model demonstrate that our approach produces explanations that are more stable, accurate, and concise than conventional attribution graphs. Moreover, through a case study on controlling cross-lingual output generation, we validated the practical effectiveness of WASD in controlling model behavior.

Large audio-language models (LALMs) can produce expressive speech, yet reliable emotion control remains elusive: conversions often miss the target affect and may degrade linguistic fidelity through refusals, hallucinations, or paraphrase. We present, to our knowledge, the first neuron-level study of emotion control in speech-generative LALMs and demonstrate that compact emotion-sensitive neurons (ESNs) are causally actionable, enabling training-free emotion steering at inference time. ESNs are identified via success-filtered activation aggregation enforcing both emotion realization and content preservation. Across three LALMs (Qwen2.5-Omni-7B, MiniCPM-o 4.5, Kimi-Audio), ESN interventions yield emotion-specific gains that generalize to unseen speakers and are supported by automatic and human evaluation. Controllability depends on selector design, mask sparsity, filtering, and intervention strength. Our results establish a mechanistic framework for training-free emotion control in speech generation.

Large language models (LLMs) are increasingly used as automated judges and synthetic labelers, especially in low-label settings. Yet these systems are stochastic and often overconfident, which makes deployment decisions difficult when external ground truth is limited. We propose a practical calibration protocol based on controlled input interventions: if noise severity increases, task performance should exhibit a statistically significant deterioration trend. We operationalize this with a slope-based hypothesis test over repeated trials, using signal-to-noise-ratio (SNR) perturbations for tabular data and lexical perturbations for text data. Across UCI tabular benchmarks and four text classification datasets, we find clear modality-dependent behavior. Our results reveal a modality gap: while text-based judges degrade predictably, the majority of tabular datasets show a lack of statistically significant performance deterioration even under significant signal-to-noise reduction. Interestingly we find that model performance is lower on datasets that are insensitive to noise interventions. We present a reproducible methodology and reporting protocol for robust LLM-judge calibration under distribution shift.

In schema-guided reasoning (SGR) pipelines, LLMs produce explicit intermediate structures -- rubrics, checklists, or verification queries -- before committing to a final decision. SGR is increasingly adopted because it promises controllability: practitioners expect to inspect, edit, and override these structures to steer the outcome. But does the promise hold? We introduce a causal evaluation protocol to measure it: by selecting tasks where a deterministic function maps intermediate structures to decisions, every controlled edit implies a unique correct output. Across 12 models and 4 benchmarks, models appear self-consistent with their own intermediate structures but fail to update predictions after intervention -- revealing that apparent faithfulness is fragile once the intermediate structure changes. When derivation of the final decision from the structure is delegated to an external tool, this fragility largely disappears; stronger prompting yields only limited improvements, while preference optimization substantially improves intervention faithfulness. Overall, intermediate structures in schema-guided pipelines function as influential context rather than stable causal mediators.

We train nine sparse autoencoders (SAEs) on the residual stream of Qwen 3.5-35B-A3B, a 35-billion-parameter Mixture-of-Experts model with a hybrid GatedDeltaNet/attention architecture, and use them to identify and steer five agentic behavioral traits. Our method trains linear probes on SAE latent activations, then projects the probe weights back through the SAE decoder to obtain continuous steering vectors in the model's native activation space. This bypasses the SAE's top-k discretization, enabling fine-grained behavioral intervention at inference time with no retraining. Across 1,800 agent rollouts (50 scenarios times 36 conditions), we find that autonomy steering at multiplier 2 achieves Cohen's d = 1.01 (p < 0.0001), shifting the model from asking the user for help 78% of the time to proactively executing code and searching the web. Cross-trait analysis, however, reveals that all five steering vectors primarily modulate a single dominant agency axis (the disposition to act independently versus defer to the user), with trait specific effects appearing only as secondary modulations in tool-type composition and dose-response shape. The tool-use vector steers behavior (d = 0.39); the risk-calibration vector produces only suppression. We additionally show that steering only during autoregressive decoding has zero effect (p > 0.35), providing causal evidence that behavioral commitments are computed during prefill in GatedDeltaNet architectures.

Multi-stream transformer architectures have recently been proposed as a promising direction for managing representation collapse and the vanishing gradient problem for residual connections, yet their internal mechanisms remain unexplored. In particular, the recently introduced Manifold-Constrained Hyper-Connections (mHC) architecture posits multiple residual streams with constrained interaction, but lacks in-depth mechanistic analysis. We present the first open-source mHC language model (https://huggingface.co/wgpeng/mhc-780m) and analyze the multiple-stream architecture with a suite of representation-level metrics and causal interventions to probe how parallel streams encode and utilize information. Specifically, we introduce a systematic stream ablation-and-rescue framework that enables direct causal comparison of residual streams during inference. Through targeted pairwise interventions and controlled recovery experiments, we distinguish functional redundancy from asymmetric utilization and reveal how information is distributed across streams beyond what is observable from representational similarity alone.

A language model can give the wrong answer even when the correct answer is decodable from its intermediate states. To study this gap between decodability and selection, we distinguish read from write at the first answer token. Read asks whether the gold token can be decoded from intermediate residual states under same-relation decoy controls. Write asks whether the final readout ranks that token first among content tokens. Under three different readers, with a randomized-label control, a substantial fraction of failures remain readable while another content token is selected. We explain this through the selection margin at the final readout, the difference between the answer logit and the logit of its strongest alternative, which is answer support minus alternative support, and can also be split into a context-averaged baseline linked to token frequency and an item-specific term. Setting the answer support to the level typical of successful generations is sufficient to recover first-token selection for the majority of failures in most of the models we study; the original alternative remains ahead in most remaining failures under this edit, and this outcome follows directly from the readout geometry. Removing the frequency direction alone shifts selection but rarely recovers the answer. Prompt variants of the same fact that succeed supply support that transfers to failing variants through the residual stream and through late MLP outputs, with less consistent effects through late attention. First-token recovery leaves most full answers wrong, which limits the recovery achieved by these edits and separates three things that are easily conflated, decodability, recoverability, and generation.

Vision-language-action policies map camera images and natural-language instructions to a robot's motor actions. Some of these policies are designed to reason in text before acting, generating a reasoning chain and decoding actions conditioned on that chain. The works introducing this design offer the reasoning chain as an oversight interface: text a person can read and edit to correct the policy. What an edited reasoning chain does to the policy's motor actions, whether it repairs them or corrupts them, has so far been measured only in part. We measure both directions, repair and corruption, with our deterministic entity swap applied to the instruction the policy receives and to the reasoning chain it generates. A forty-task observed backdrop across all four LIBERO simulation suites reveals that the cost of corrupting the reasoning chain concentrates where language alone determines the goal. There, on LIBERO-Goal, we run the counterfactual intervention with DeepThinkVLA, chosen because its reasoning chain is exposed as plain text. The policy receives a corrupted instruction, but its reasoning chain is replaced by the one it generates when that instruction is clean. This counterfactually correct reasoning chain recovers 47.8 pp of the lost success, our pre-registered confirmatory test. Had the chain merely restated what the camera image already determines, the replacement could have changed nothing. Instead, all 10 tasks move in the predicted direction, though success falls short of the clean runs by 38.0 pp, a gap we had predicted at 5-15 pp. The reasoning chain is therefore a working control surface: text written into it moves the robot, repairing behaviour when the text is right and corrupting it when the text is wrong. Whether to expose such a control surface is a deployment tradeoff, and part of it can now be measured.

Large language models are often viewed as high-dimensional nonlinear systems and treated as black boxes. Here, we show that transformer depth dynamics admit accurate low-order linear surrogates within context. Across tasks including toxicity, irony, hate speech and sentiment, a 32-dimensional linear surrogate reproduces the layerwise sensitivity profile of GPT-2-large with near-perfect agreement, capturing how the final output shifts under additive injections at each layer. We then uncover a surprising scaling principle: for a fixed-order linear surrogate, agreement with the full model improves monotonically with model size across the GPT-2 family. This linear surrogate also enables principled multi-layer interventions that require less energy than standard heuristic schedules when applied to the full model. Together, our results reveal that as language models scale, low-order linear depth dynamics emerge within contexts, offering a systems-theoretic foundation for analyzing and controlling them.