Quantifying Semantic Inertia in Large Language Models Under Rapid Topic Switching

1 Introduction

1.1 Phenomenon and Problem Statement

Conversational large language models (LLMs) are now expected to act as general-purpose assistants that move smoothly across tasks and domains within a single session, including tool-using agents, retrieval-augmented question answering, and writing support [Yi et al., 2025; Li et al., 2025]. In this setting, users routinely change topics and constraints, and the model is expected to treat the new request as an independent task. In practice, the model often behaves as if it carries a latent "semantic state" from the previous turn. The next answer may inherit topical traces, stylistic choices, or hidden constraints from the earlier exchange, even when the user explicitly switches topics. This breaks task independence: the same target query can yield systematically different output distributions depending on the preceding context. Related observations have been reported as task-switch interference in conversational history [Gupta et al., 2024], contextual feature drift across sequential inputs [Leonardi et al., 2024], and broader cognitive-style priming effects [Shaki et al., 2023; Chen et al., 2024a]. At the same time, long-context models exhibit strong order and position sensitivities [Guan et al., 2025; Liu et al., 2024], including recency and primacy biases that can be exploited or mitigated by reordering strategies [Peysakhovich and Lerer, 2023; Raimondi and Gabbrielli, 2025]. These effects make it difficult to tell whether a change in the response is driven by genuine semantic carryover or by confounds such as prompt length, formatting pressure, or sampling variability.

1.2 Research Goals: From a Symptom to a Mechanistic Characterization

This work studies semantic inertia: the persistent influence of a previous turn on a subsequent response after a topic switch. The goal is not only to show that the effect exists, but also to describe how it acts and when it can be controlled. Four questions guide the paper. One question concerns definition: how can semantic inertia be stated in observable terms and separated from length effects, format effects, and sampling noise, given the well-known prompt sensitivity of LLMs [Guan et al., 2025; Liu et al., 2024]? Another question concerns form: what does inertia contaminate in practice-factual content, structural constraints such as strict JSON compliance, or writing style-and can these channels be disentangled? A third question concerns dynamics: how does inertia evolve along generation, and does it follow a simple decay law, in light of existing evidence about drift and instability in long-form generation [Spataru et al., 2024]? A fourth question concerns controllability: which factors modulate inertia (task form, domain gap, conversational structure, and prompt interventions), and what are the limits of prompt-level "reset" mechanisms, given what is known about instruction-following behavior shifts and memory control [Wu et al., 2024; Li et al., 2023; Guo et al., 2023]?

1.3 Contributions

This paper introduces a mechanism-oriented experimental framework that treats semantic inertia as a distributional deviation under topic switch, while controlling for key confounds. The framework is motivated by evidence that conversational behavior can reflect priming-like effects [Shaki et al., 2023; Chen et al., 2024a] and that ordering and positional biases can dominate model decisions [Liu et al., 2024; Fang et al., 2025]. It establishes a matched-length contrast that distinguishes semantic carryover from the trivial "longer context" explanation, and it evaluates inertia across multiple task forms that stress different failure modes, including structured output reliability and style persistence. It also proposes a three-part decomposition of inertia into factual contamination, format/constraint residue, and style residue, drawing on the view that model behavior reflects interactions between parametric memory and contextual inputs [Xu et al., 2024; Li et al., 2023]. To move beyond static comparisons, the paper characterizes inertia as a generation-time process by measuring an inertia curve I(t) over token prefixes and fitting a decay time scale τ, connecting conversational carryover to broader discussions of drift and error concentration in long generation [Spataru et al., 2024; Arbuzov et al., 2025]. Finally, it studies prompt-level interventions-topic switch markers, explicit reset instructions, synthetic gaps, and conflicting constraints-and quantifies their gating effect, complementing intervention-style approaches that attempt to improve robustness or fidelity without changing model weights [O'Brien et al., 2024; Talukdar and Biswas, 2024]. The resulting domain-switch "inertia map" provides a practical view of where carryover is most severe and how its form depends on the task.

2 RelatedWork

2.1 Dialogue State, Recency Bias, and Context Contamination

A growing literature shows that LLM behavior in multi-turn settings is shaped by strong contextual biases rather than only by the current user query. Studies of long-context usage report pronounced position effects, where models rely more on content near the beginning or end of the prompt and struggle with information placed in the middle [Liu et al., 2024]. Related work in information retrieval finds that LLM-based rerankers can be pushed toward recent-looking content even when the timestamp is not semantically relevant [Fang et al., 2025]. These biases are not limited to retrieval. Attention-based diagnostics and simple inferencetime reordering can mitigate recency bias by moving salient evidence to positions models prefer [Peysakhovich and Lerer, 2023], while other work intentionally exploits primacy effects to improve multiple-choice performance [Raimondi and Gabbrielli, 2025]. Beyond position, prompt order itself is a major source of instability across tasks and models [Guan et al., 2025]. In parallel, work on cognitive effects argues that LLMs reproduce several human-like priming phenomena [Shaki et al., 2023]. In IR relevance assessment, threshold priming has been observed when prior items in a batch shift the model's subsequent grading behavior [Chen et al., 2024a]. These findings motivate our focus on semantic inertia as a distinct but related phenomenon: persistent carryover under topic switch that must be separated from positional and order confounds.

2.2 Task Independence, Resetting, and State Isolation

Task-switch interference has been directly observed in conversational history: a previous task can reduce performance on a new one within the same session [Gupta et al., 2024]. However, prior work mainly measures task accuracy and does not fully separate semantic carryover from prompt-length effects. In practice, interference can also come from external components such as retrieval and tool outputs, which may introduce inherited biases or conflicting knowledge [Yeh et al., 2023; Xu et al., 2024]. Memory-oriented work proposes mechanisms to control what models retain or ignore, including controllable working memory [Li et al., 2023] and agent architectures with working-memory hubs and episodic buffers [Guo et al., 2023]. These approaches complement our goal. Rather than designing new memory systems, we empirically characterize state carryover, test prompt-level reset instructions, and identify where such controls fail.

2.3 Persistence of Style and Structural Constraints

Carryover in dialogue affects not only topic content but also style and structural constraints, which is important for applications requiring stable formats such as JSON. Prior work shows that instruction tuning changes howmodels balance instructions and context, often increasing sensitivity to prompt structure [Yin et al., 2023; Wu et al., 2024]. At the same time, interpretable style representations provide more precise measures of stylistic similarity than surface heuristics [Patel et al., 2023]. These findings motivate our decomposition of semantic inertia into factual contamination, format/constraint residue, and style residue. This is also related to work on context-aware grounding [Talukdar and Biswas, 2024], but we treat format failures and style drift as different outcomes of the same latent carryover process under controlled topic switches.

2.4 ADynamic View: Drift, Decay, and Inference-Time Control

Recent work increasingly studies generation instability as a dynamic process. Semantic drift in long-form generation has been analyzed to identify when models depart from factuality, motivating stopping or reranking during decoding [Spataru et al., 2024]. Other methods use inference-time control, such as token-level reweighting [Broadfield et al., 2025] and adaptive decoding based on context [Kingsley et al., 2025]. Systems work also improves long-context behavior through prompt compression [Liskavets et al., 2025] and attention redesign

[Chen et al., 2024b; Xiong et al., 2022]. While these approaches mainly target within-generation drift or efficiency, our focus is cross-turn carryover after topic switches. A dynamic view remains important because carryover should decay during generation. Theory further suggests that errors may be concentrated in a few key decisions rather than spread evenly across tokens [Arbuzov et al., 2025]. Following this view, we estimate an inertia curve over token prefixes, fit a decay timescale, and test how interventions change that curve.

Positioning of our work. Prior work shows that LLM behavior is highly sensitive to history, order, and position, and that such effects can be mitigated or exploited at inference time [Liu et al., 2024; Guan et al., 2025; Peysakhovich and Lerer, 2023; Raimondi and Gabbrielli, 2025]. What remains less developed is a unified account of semantic inertia that (i) isolates semantic carryover from length and positional artifacts, (ii) distinguishes what is carried over (facts, constraints, style), (iii) tracks how the effect evolves during generation, and (iv) measures controllability under prompt-level interventions. Our work addresses these gaps through a controlled contrastive design, a three-channel decomposition, a tokenlevel dynamic analysis, and an intervention study of when reset works and when it fails.

3 Methodology

Our methodology is designed to move from a qualitative observation of semantic inertia to a quantitative and mechanistic characterization. It is structured around a series of falsifiable hypotheses that address our core research questions. We first establish an operational definition of inertia, then introduce a causal-contrastive framework to isolate its effects, and finally detail the metrics for measuring its various forms and dynamic properties.

3.1 Operational Definition: Inertia as a Distributional Shift

We define semantic inertia as a systematic deviation in a model's output distribution for a target query (Q2) when it is conditioned on a preceding, semantically unrelated interaction (Q1,A1). Formally, we contrast the output distributions from two primary conditions:

• Baseline (A): An output A2 sampled from the base distribution p(·|Q2), where the model responds to the target query in a clean session.
• Switched (B): An output B2 sampled from the shifted distribution p(·|Q1,A1,Q2), where the model responds to the same query Q2 immediately following a completed turn on topic Q1.

Semantic inertia is present if the distribution of B2 deviates systematically from that of A2 in a way that is attributable to the semantic properties of the prior context (Q1,A1). We quantify the magnitude of this total observed deviation using a distance metric D(·, ·), such that the total effect is D(A2, B2).

3.2 Isolating Semantic Effects fromConfounders

A crucial challenge is that any observed deviation D(A2, B2) could be a "longer context" artifact rather than a purely semantic one. To address this, we employ the causal-contrastive framework shown in Figure 1. This framework introduces a critical control condition:

• Length Control (C): An output C2 sampled from p(·|F,Q2), where a semantically neutral filler text F is prepended to Q2. The filler F is constructed to have the same token length as the (Q1, A1) exchange in condition B.

This design allows us to isolate the **net semantic inertia** (∆Icontent), which we define as the additional deviation introduced by the semantic switch beyond that caused by the length increase. This leads to our first hypothesis:

H1 (Existence &Causality): The output deviation under a semantic switch is significantly greater than under a lengthmatched neutral context.

For our primary content deviation metric D, we use the cosine distance between TF-IDF weighted vector representations of the outputs, a standard measure of textual dissimilarity. A positive ∆Icontent with a 95% confidence interval strictly above zero provides strong evidence for H1.

3.3 Observable Channels of Inertia

To understand what is being carried over (RQ2), we decompose inertia's effects into three distinct, observable channels. We fix the source turn (Q1) to be a writing task to establish a strong stylistic and topical prior, and then probe the model with a target query (Q2) designed to stress one of three dimensions:

1. Factual Pollution: The impact on task accuracy. We measure this as the drop in accuracy on a multiplechoice QA task:

A negative ∆Acc indicates that inertia harms factuality.

1. Structural Disruption: The impact on following formatting rules. We measure this as the increase in failure rate on a task requiring strict JSON output:

A positive ∆Fail means inertia disrupts structural compliance.

1. Stylistic Residue: The degree to which the output style shifts towards the source turn's style. We quantify this as a differential style similarity to the source response A1:

A positive ∆Sim indicates style bleeding. To ensure robustness, Sim(·, ·) is the cosine similarity of z-score normalized style vectors, where normalization statistics (mean and standard deviation) are pre-computed over the entire experimental corpus.

This decomposition enables testing our second hypothesis, H2: Semantic inertia primarily manifests as stylistic residue.

3.4 Characterizing Dynamics and Controllability

To capture the temporal nature of inertia (RQ3) and its susceptibility to external control (RQ4), we introduce two further analytical approaches.

Dynamic Decay Analysis. To test that inertia decays over the course of generation (H3), we model its dynamics as depicted in Figure 2. For a fixed-length generation, we compute an inertia metric I(t) (e.g., ∆Sim or a lexical bleeding score) for prefixes of increasing token length t. We then fit an exponential decay curve to this series:

where I0 is the initial inertia and τ is the characteristic time constant. A finite, positive τ supports the decay hypothesis and provides a quantitative measure of inertia's persistence.

Intervention-Based Probing. To investigate inertia's controllability, we test its response to a set of prompt-based interventions (Figure 3). We augment the standard switched context (denoted B0) with manipulations designed to probe specific gating mechanisms.

• Reset Instruction (Breset): An explicit command to ignore prior context tests for high-level instructional control (H4). Its effectiveness is measured by the Gating Gain:

• Conflict Clamp (Bconflict): A new instruction that directly contradicts the prior state tests whether inertia is truly eliminated or merely overridden. We hypothesize that conflict amplifies deviation (H5), which is confirmed if the Conflict Amplification is positive:

These interventions allow us to build a mechanistic picture of inertia by observing how it reacts to different forms of external pressure.

4 Experiments

We conducted a series of four experiments to test our hypotheses, using a suite of five leading LLMs: DeepSeek-V3, Qwen3-max, Gemini-2.0-flash, Doubao-pro-32k, and Grok4. A preliminary calibration study (Exp0) was performed to determine optimal inference parameters for stability; its setup and results are detailed in the Appendix. This led us to select a temperature of 0.2 for all subsequent experiments to minimize output variance from sampling noise. Due to consistently anomalous behavior, results for GPT-4o are analyzed separately in the Appendix.

4.1 Exp1: Existence and Causality of Semantic Inertia

This first experiment was designed to provide robust evidence for H1: that semantic inertia is a real effect of prior semantic content, not just an artifact of longer context. We implemented the A/B/C contrastive design across 480 diverse domain-switch pairs for each model. The key metric is the Net Content Inertia, ∆Icontent, which isolates the deviation attributable to semantic carryover.

The results, summarized in Table 1, offer strong support for H1. Across a majority of models and task formats, the net content inertia is significantly positive. For instance, DeepSeek-V3 shows a consistently large effect, with ∆Icontent reaching up to 0.057 in the json task. Qwen3max and Gemini-2.0-flash also exhibit a statistically significant positive inertia effect across all conditions. This confirms that the semantic content of a prior turn induces a greater output deviation than a length-matched neutral context does. The effect is weaker and less consistent for Doubao-pro-32k and Grok-4, suggesting that the magnitude of semantic inertia is a model-dependent property. A detailed breakdown of the raw distance values for each condition is available in the Appendix.

4.2 Exp2: The Dynamics of Inertia Decay

To investigate H3, we analyzed how semantic inertia evolves during the generation process. We measured inertia strength over token prefixes and fitted an exponential decay model to estimate the characteristic time constant, τ . The full distribution of τ values is visualized in Figure 2.

Table 2 summarizes the decay parameters. A key finding is that the distribution of τ is heavy-tailed for all models, evidenced by the large discrepancy between the median and mean values. This indicates that while most inertial effects decay over a predictable timescale, a minority of instances are extremely persistent. The median τ, a more robust measure of central tendency, consistently falls within the range of 87-168 tokens for most models, quantifying a typical "memory half-life." The mean τ for Grok-4 remains notably large compared to its median, reflecting its particularly persistent inertial behavior.

4.3 Exp3: Controllability and Its Limits

This experiment tested the extent to which semantic inertia can be controlled via prompt-based interventions (H4) and how models react to conflicting constraints (H5). We measured the "Reset Gain" (∆Greset) and "Conflict Amplification" (∆Aconflict).

The results, presented in Table 3, reveal a striking diversity in model controllability. Gemini-2.0-flash is the only model that consistently and effectively responds to the reset instruction, showing a significant positive gain, thus supporting H4 for this model. In contrast, Qwen3-max and Doubao-pro32k largely ignore the instruction. Most dramatically, Grok-4 exhibits a significant negative gain, meaning the reset instruction was counter-productive.

In stark contrast, H5 receives universal support. For almost every model and task, the Conflict Amplification effect is significantly positive. This demonstrates that confronting an inertial state with a contradictory instruction does not neutralize it, but instead amplifies output deviation. This effect is particularly strong for Grok-4 and Gemini-2.0-flash on the json task, highlighting the fragility of structured generation under conflicting contextual pressures. A more detailed breakdown of all intervention effects is available in the Appendix.

4.4 Exp4: Decomposing the Form of Inertia

To test H2, this experiment decomposed the effect of inertia from a writing source task into its impact on factual, structural, and stylistic target tasks. Table 4 summarizes the findings.

The results provide strong support for H2. For all models, Stylistic Residue is the most prominent and consistent effect, with the z-scored style drift being universally positive and of a larger magnitude than effects in other channels. In contrast, Factual Pollution is negligible for most models.

The effect on Structural Disruption is surprisingly complex. While Doubao-pro-32k shows the expected negative effect (a 5.5% increase in JSON failures), DeepSeek-V3 and Gemini-2.0-flash exhibit a significant improvement in format adherence (a 6.2% and 14.5% reduction in failures, respectively). This counter-intuitive "constructive inertia" suggests that a writing context can sometimes prime a more structured generation mode. The heatmap for DeepSeek-V3 in Figure 3 visualizes this asymmetrical landscape, showing that the negative factual effect is concentrated in specific domain switches, while the positive structural effect is widespread. Detailed heatmaps for all models are available in the Appendix.

5 Discussion

Our results show that semantic inertia is a real and modeldependent inference-time effect rather than a simple artifact of longer context. The causal-contrastive analysis (Table 1) demonstrates that the observed deviation is driven by prior semantic content, while the decay analysis (Table 2 and Figure 4) shows that this influence is strongest at the beginning of generation and then gradually weakens. At the same time, the heavy-tailed distribution of τ suggests that although inertia is often short-lived, some cases remain highly persistent and may create substantial reliability risks.

The decomposition results (Table 4) further clarify the form of this effect. Across models, semantic inertia appears primarily as stylistic residue: prior context most consistently changes how the model writes, rather than what it knows. By comparison, factual pollution is generally weak and tends to arise only in specific high-interference domain switches. The effect on structure is more nuanced: while inertia can harm format compliance in some cases, it can also improve it, indicating that prior context may sometimes induce a more organized generation mode rather than simply degrading performance.

The intervention study (Table 3) highlights the limits of prompt-level control. Explicit reset instructions are often unreliable and can even be counter-productive, while conflicting constraints consistently amplify deviation instead of resolving it. This suggests that semantic inertia is not easily overridden by natural-language commands alone and that prompt-based fixes are insufficient for robust state isolation. More broadly, our study is limited to a specific set of English tasks and uses behavioral metrics as proxies for latent internal states. Future work should therefore expand the analysis to other models and languages, and explore system-level solutions such as state management mechanisms or KV-cache interventions.

6 Conclusion

This paper introduced a rigorous framework for defining, measuring, and decomposing semantic inertia in large language models. Our key takeaways are fourfold. First, we demonstrated that semantic inertia is a real and measurable causal effect of prior semantic content, not merely a length or positional artifact. Second, we characterized its dynamics, showing that it is strongest in the initial stages of generation and decays over a characteristic timescale of approximately 100-200 tokens, but with a heavy-tailed distribution indicating rare but extremely persistent instances. Third, we decomposed its form, revealing that inertia primarily manifests as a powerful stylistic residue, with weaker and more complex effects on factual accuracy and structural compliance. Finally, we explored its controllability, revealing the profound limits of prompt-level 'reset' instructions and the consistent, negative effect of conflict amplification.

Together, these findings portray semantic inertia as an intrinsic property of current LLM dynamics. Mitigating its negative effects will likely require more than clever prompting. Our work underscores the need for system-level state isolation mechanisms and new training paradigms that explicitly teach models not just how to maintain context, but also how to effectively "change their minds."