Hallucination Reduction with CASAL: Contrastive Activation Steering for Amortized Learning¶

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

Contrastive activation steering reduces hallucination via linear knowledge representations

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

CASAL (Contrastive Activation Steering for Amortized Learning) is a lightweight training method that bakes inference-time activation steering into model weights by training only a single transformer layer with a pure representation loss. It reduces hallucination rates by ~30–40% on short-form QA benchmarks while being ~30× more compute-efficient and ~20× more data-efficient than LoRA-based baselines. The method is modality-agnostic, working on both text-only and vision-language models, and on both dense and MoE architectures.

Problem¶

Prior activation-steering methods that reduce LLM hallucination require real-time inference-time intervention, making them impractical for scalable deployment. Fine-tuning baselines (SFT, DPO, GRPO) are data- and compute-expensive. No prior method amortized the steering process into model weights via a representation-only training objective.

Method¶

CASAL proceeds in three stages:

Knowledge boundary probing: For each training query $x$, sample $k=10$ completions; classify as known ($D_k$) if $\geq 7$ are correct, unknown ($D_u$) if $\leq 3$ are correct (threshold $\tau=7$).
Steering vector construction: Extract residual stream activations $a^{L^*}(x)$ at a target layer $L^*$. Compute per-class mean activations $\bar{a}^{L^*}_k$ and $\bar{a}^{L^*}_u$, then form contrastive steering vectors $v^{L^*}_u = \bar{a}^{L^*}_u - \bar{a}^{L^*}_k$ and $v^{L^*}_k = \bar{a}^{L^*}_k - \bar{a}^{L^*}_u$. Target activations are $t^{L^*}_u(x) = a^{L^*}(x) + \alpha \cdot v^{L^*}_u$ and $t^{L^*}_k(x) = a^{L^*}(x) + \alpha \cdot v^{L^*}_k$.
CASAL training: Train a one-layer network $M_\text{train}$ initialized from $W^{L^*}_\text{original}$ using the sole MSE objective: $$\mathcal{L} = \mathbb{E}_{x \in D_u}\|t^{L^*}_u(x) - a^{L^*}(x)\|^2 + \mathbb{E}_{x \in D_k}\|t^{L^*}_k(x) - a^{L^*}(x)\|^2$$ The trained weights $W^{L^*}_\text{CASAL}$ replace layer $L^*$ in the original frozen model. No cross-entropy loss is used.

Because the loss is local to $L^*$, neither forward nor backward passes traverse the full model depth.

Key Contributions¶

Introduces the first training method that uses representation loss exclusively (no cross-entropy) to amortize activation steering into model weights.
~30–40% hallucination reduction on TriviaQA, PopQA, and EntityQA with as few as 640 training examples.
~30× fewer FLOPs per token and ~20× less training data than LoRA-based SFT/DPO/GRPO.
OOD generalization: trained on TriviaQA, hallucination rate on unseen EntityQA drops from 50.7% to 11.7%.
Extends to VLMs (Qwen2.5-VL-7B-Instruct): 38.74% hallucination reduction on WorldCuisines-VQA.
First steering-based training framework demonstrated on both dense and MoE (OLMoE) architectures.

Results¶

Text QA (in-distribution): CASAL achieves refusal rates of 7.29% (TriviaQA) and 6.84% (EntityQA) vs. baselines ranging 10–18%, while accuracy on known queries is preserved (95.34% / 89.90%).
General capability: CASAL matches baseline on MMLU (68.04 vs. 68.01), GSM8K (77.02 vs. 77.48), GPQA (33.18 vs. 33.31), and MT-Bench (7.57 vs. 7.38) — no capability degradation.
OOD transfer (train TriviaQA → test EntityQA): hallucination rate 11.72% vs. baseline 50.74%; accuracy 95.77% vs. baseline 95.82%.
VLM (WorldCuisines-VQA): hallucination rate 33.34% vs. baseline 72.35%; accuracy 90.36% vs. 76.72%.
VLM (Landmark-VQA): hallucination rate 31.25% vs. baseline 75.78%; accuracy 99% vs. 90.80%.
MoE (OLMoE): hallucination rate drops by 42.9%; known-query accuracy unchanged.
Compute: >30× fewer FLOPs than LoRA; data: 640 CASAL examples match 12,800 SFT/DPO/GRPO examples.
Silhouette score of known/unknown cluster separation correlates strongly with hallucination reduction ($R^2 = 0.945$, logistic fit).

Limitations¶

Evaluated only on short-form QA benchmarks; generalization to long-form generation tasks is unverified.
Requires $k=10$ sampled completions per training query for knowledge probing, adding upstream inference cost (shared with baseline labeling pipelines).
Target layer $L^*$ selection requires a hyperparameter search or heuristic; sensitivity analysis is deferred to the appendix.
VLM experiments use only two visual QA datasets (WorldCuisines, Landmark); broader multimodal benchmarks (e.g., MMBench, POPE) are not reported.
The method trains a single-layer submodule; whether this suffices for models with highly distributed knowledge representations (very large MoE with many experts) is not fully characterized.

Relevance to Vision-Language Models¶

CASAL directly applies to VLMs by treating the multimodal transformer's residual stream at a target layer identically to the text-only case, demonstrating that knowledge boundary representations are present and steerable in models like Qwen2.5-VL-7B-Instruct. This is significant for VLM researchers because hallucination (confabulating object attributes, non-existent entities, incorrect visual facts) is a first-order reliability problem in deployed VLMs, and CASAL offers a parameter-efficient, training-time solution that does not require reward models or preference data. The modality-agnostic design means the technique could extend to other VLM architectures without architectural modifications. Combined with its OOD generalization, CASAL represents a practical complement to contrastive decoding and RLHF-based VLM alignment methods.