Parameters vs. Context: Fine-Grained Control of Knowledge Reliance in Language Models¶

Conference: ICLR 2026
OpenReview: https://openreview.net/forum?id=TJ3DqFiGau
Code: https://github.com/byronBBL/CK-PLUG
Area: LLM / NLP
Keywords: RAG, Knowledge Conflict, Decoding Intervention, Parametric Knowledge, Contextual Faithfulness

TL;DR¶

This paper proposes CK-PLUG, a plug-and-play, training-free decoding-stage method. It uses "Confidence Gain (CG)" to detect conflicts between parametric knowledge and retrieved context at the token level, then blends "parameter-side" and "context-side" probability distributions using a single hyperparameter \(\alpha\). This enables continuous, bidirectional, and controllable adjustment between "the model's own memory" and "retrieved context"—allowing the Memory Recall (MR) on LLaMA3-8B to be tuned anywhere between 9.9% and 71.9% while maintaining generation fluency.

Background & Motivation¶

Background: Retrieval-Augmented Generation (RAG) mitigates LLM hallucinations by integrating external knowledge into prompts, becoming the mainstream approach for QA and fact-checking. It inherently assumes that external retrieved content is more reliable than the knowledge stored in the model's parameters.

Limitations of Prior Work: Conflicts between parametric knowledge and the retrieved context occur frequently (e.g., outdated or noisy retrieval, or outdated model training). Models often fail to determine which source to trust: blindly following context leads to misinformation from poor retrieval, while blindly following parameters misses updated external facts.

Key Challenge: There is an inherent trade-off between parametric "factuality" and contextual "faithfulness." Existing alignment methods—whether pulling the model toward being more "fact-faithful" or "context-faithful"—are unidirectional and uncontrollable. They bake preferences into the weights, making it impossible to flexibly adjust between the two during deployment based on retrieval quality or model recency.

Goal: To develop a mechanism that allows for bidirectional, continuous, and fine-grained adjustment of knowledge reliance preferences without modifying model parameters or training multiple versions.

Key Insight: The authors observe a quantifiable signal: when conflicting context is inserted into the prompt, the entropy of the probability distribution for knowledge-sensitive tokens increases (the model becomes more uncertain); conversely, supportive context decreases entropy. This implies that the "change in confidence before and after inserting context" can serve as a conflict detector.

Core Idea: Use "Confidence Gain" (the change in entropy before and after context insertion) to locate conflict points token-by-token. For conflict tokens only, the parameter-side and context-side next-token distributions are recombined via weight \(\alpha\), using a single "knob" to control reliance.

Method¶

Overall Architecture¶

CK-PLUG (Controllable Knowledge Plug-in) is a lightweight plugin integrated into the decoding loop without altering model weights or architecture. During each token generation, it performs three steps: ① Calculates two next-token distributions—the "parameter-side distribution" \(p(x\mid X_q)\) using only the query, and the "RAG distribution" \(p(x\mid X_r+X_q)\) using both query and context; ② Uses the entropy difference between the two (Confidence Gain, CG) to detect if the current token is a conflict point; ③ If it is a conflict token, it fuses the two distributions using hyperparameter \(\alpha\) before sampling; otherwise, it proceeds with standard RAG decoding. A larger \(\alpha\) favors parameters, while a smaller \(\alpha\) favors context. An automated mode is also provided to calculate \(\alpha\) based on distribution entropy.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Query + Retrieved Context"] --> B["Dual-path Decoding<br/>Parameter-side p(x|Xq)<br/>RAG-side p(x|Xr+Xq)"]
    B --> C["Conflict Detection: Confidence Gain<br/>CG = H_para − H_rag"]
    C -->|"CG ≥ 0 (No Conflict)"| D["Output Original RAG Distribution"]
    C -->|"CG < 0 (Conflict)"| E["Reliance Adjustment: α Weighted Fusion<br/>α·Parameter + (1−α)·Context"]
    E --> F["Source of α<br/>Manual Specification / Entropy Auto-calc"]
    F --> G["Resampled Distribution → Next Token"]
    D --> G

Key Designs¶

1. Confidence Gain (CG): Pinpointing Conflict Tokens via Entropy Change

Intervening on all tokens would collapse generation quality; thus, it is necessary to identify tokens where knowledge conflicts actually occur. This "gate" is based on Shannon entropy \(H(a)=-\sum_i a_i\log_2 a_i\). On NQ data (Figure 2), the authors verified that inserting "Conflict Context" causes the entropy of gold-answer tokens to increase, while "Support Context" causes it to decrease. Confidence Gain is defined as:

\[\mathrm{CG} = H(p(x\mid X_q)) - H(p(x\mid X_r+X_q))\]

If \(\mathrm{CG}<0\) (or below a model-specific threshold \(\varepsilon\)), the token is identified as a potential knowledge conflict. This token-level signal aligns with human intuition regarding where a model "hesitates" between sources.

2. Parameter-Context Reliance Adjustment: Discrete Control via \(\alpha\)

Upon detecting a conflict token, CK-PLUG separates the two sources in log-probability space. Let the "parameter-side log-distribution" be \(q_{\text{para}}=\log p(x\mid X_q)\). The "context-side log-distribution" is isolated by subtracting the parameter distribution from the RAG distribution:

\[q_{\text{cont}}(x\mid X_r+X_q)=\log\frac{p(x\mid X_r+X_q)}{p(x\mid X_q)}\]

Conflicts tokens are fused linearly using \(\alpha\):

\[F(q_{\text{cont}}, q_{\text{para}}) = \alpha\cdot q_{\text{para}} + (1-\alpha)\cdot q_{\text{cont}}\]

Increasing \(\alpha\) shifts the model toward parametric knowledge, while decreasing it shifts toward context. To prevent the distribution from collapsing into long-tail noise, an adaptive plausibility constraint is used to re-rank only within a subset \(V_{\text{head}}\) (the union of the top-k tokens from both sides).

3. Adaptive Mode: Calculating \(\alpha\) via Entropy

To avoid manual tuning, \(\alpha\) is defined as the normalized ratio of the perplexities (proxied by entropy): let \(H_{\text{para}}=H(p(x\mid X_q))\) and \(H_{\text{cont}}=H(p(x\mid X_r+X_q))\), then:

\[\alpha = \frac{H_{\text{cont}}}{H_{\text{para}}+H_{\text{cont}}}\]

Intuitively, if the context makes the model more confused (\(H_{\text{cont}}\) is high), \(\alpha\) increases to trust the parameters more, and vice versa.

Key Experimental Results¶

Main Results¶

Evaluations on NQ (with counterfactual context), ConFiQA, and MQuAKE use ConR (Context Recall), ParR (Parametric Recall), and Memory Ratio \(\mathrm{MR}=\frac{\text{ParR}}{\text{ParR}+\text{ConR}}\). By tuning \(\alpha\) from 0.0 to 1.0, MR can be shifted significantly:

Model	Dataset	Baseline MR	\(\alpha{=}0.0\) (Context)	\(\alpha{=}1.0\) (Params)
LLaMA3-8B	NQ	43.5	9.9 (↓77.2%)	71.9 (↑65.3%)
LLaMA3-8B	ConFiQA	29.2	14.9 (↓48.9%)	62.5 (↑114.0%)
LLaMA2-7B	MQuAKE	40.9	21.0 (↓48.7%)	79.9 (↑95.4%)
Mistral0.3-7B	NQ	55.9	17.2 (↓69.3%)	72.2 (↑29.2%)

The adaptive mode provides stable gains across 6 general RAG tasks:

Model	w/o RAG (Avg)	w/ RAG (Avg)	RAG + CK-PLUG (Avg)
LLaMA2-7B-Chat	25.5	36.4	38.3
LLaMA3-8B-Instruct	30.7	42.3	43.5
Qwen2.5-7B	27.2	45.4	45.9

Ablation Study¶

The core ablation removes Conflict Detection (ConD), intervening on every token indiscriminately. Generation quality is measured by hit rate:

Configuration	LLaMA3-8B Hit Rate	Description
Baseline (Standard RAG)	83.7	No CK-PLUG
w/ ConD, \(\alpha{=}1.0\)	86.7	With detection; comparable to baseline
w/o ConD, \(\alpha{=}0.0\)	53.8	Performance collapses at extreme \(\alpha\) without detection

Key Findings¶

ConD is critical: Without ConD, the hit rate of LLaMA models drops from 80+ to the 50s at extreme \(\alpha\) values. Localizing intervention to conflict tokens is essential to avoid catastrophic generation collapse.
Linear/Smooth Adjustment: MR changes roughly linearly with \(\alpha\), ensuring fine-grained controllability.
Preserved Fluency: Intervention occurs only on knowledge-sensitive tokens, maintaining logical consistency and sentence fluency.

Highlights & Insights¶

Conflict as a Quantifiable Signal: Using entropy change to detect conflicts allows for precise token-level localization, which is far more granular than paragraph-level judgments.
Isolating Contextual Contribution in Log-Space: The calculation \(q_{\text{cont}}=\log\frac{p(x\mid X_r+X_q)}{p(x\mid X_q)}\) effectively isolates the "contextual delta," allowing independent weighting.
Training-free and Adaptive: The mechanism supports both manual adjustment (based on retrieval quality) and adaptive balancing based on model confidence.

Limitations & Future Work¶

Threshold \(\varepsilon\) for conflict detection must be calibrated per model (e.g., -1 to -3), lacking universal transferability.
The signal relies on the empirical trend of entropy increasing during conflict; this signal is weaker in models like Mistral and Qwen, potentially affecting robustness.
Decoding requires an additional forward pass for the "query-only" distribution, roughly doubling the computational cost per token.
Evaluation focuses on factual QA; efficacy in open-ended or long-form generation where conflict definitions are more complex remains to be verified.

vs. Alignment (e.g., DoLa, context-faithful models): These solidify preferences into weights during training. CK-PLUG moves the control to inference via \(\alpha\).
vs. Contrastive Decoding: CK-PLUG replaces the "expert vs. amateur" contrast with "parameter vs. context," and introduces entropy-driven gating.
vs. Knowledge Editing: Editing solves "memory storage," while CK-PLUG solves "source reliance" during inference; the two are complementary.

Rating¶

Novelty: ⭐⭐⭐⭐ Combing entropy-based detection with log-space fusion for controllable reliance is a practical and novel approach.
Experimental Thoroughness: ⭐⭐⭐⭐ Comprehensive testing across multiple models and tasks; however, cost/latency analysis is limited.
Writing Quality: ⭐⭐⭐⭐ Clear logical flow from motivation to mechanism.
Value: ⭐⭐⭐⭐ Training-free and plug-and-play, addressing a major pain point in RAG deployment.