When a language model answers a question from retrieved documents, irrelevant passages do two kinds of damage. They pull the answer toward the wrong entity, and they leave the model's confidence high while they do it. A zero-shot reader handed an unrelated document will name a wrong entity with probability 0.98 (Improving Zero-shot Reader by Reducing Distractions from Irrelevant Documents in Open-Domain Question Answering). Current fixes treat these as separate problems. Some filter the text before the model reads it. Others recalibrate confidence afterwards. We propose a single mechanism aimed at both. The retrieved passages would be forced to reach the question-and-answer tokens only through a small set of learned "bottleneck tokens", a narrow channel of a few dozen vectors borrowed from multimodal video/audio fusion. The model's confidence would then be read directly off those tokens.
The bet is this. If everything the context contributes has to pass through $B$ vectors, those vectors are the natural place to measure how much usable evidence actually got through. That measurement should stay calibrated as the context fills with noise, which the final output probabilities do not. There is a known risk. Compressing a representation too hard destroys the fine-grained information that uncertainty estimates depend on (Calibration Bottleneck: Over-compressed Representations are Less Calibratable). So the core scientific question is where, as the channel width $B$ varies, noise suppression and calibratability stop pulling against each other.
Core Claim
Inside a decoder language model, route all communication between retrieved passages and the question/answer stream through $B$ learnable latent tokens. Introduce this routing at a middle layer, and train a correctness head on those tokens with a proper scoring rule, meaning a loss that is minimized only by reporting true probabilities. We claim this yields two things. First, confidence estimates degrade less as distractor passages are added than logit-based or full-hidden-state probe estimates on an equally tuned standard model. Second, calibratability as a function of $B$ is U-shaped. Very small $B$ over-compresses, echoing the layer-wise U-curve seen in vision networks. Very large $B$ readmits the noise.
We are not claiming that a latent bottleneck beats text filtering on accuracy. The claim concerns what the bottleneck tokens reveal about reliability under noise. It stands or falls on a direct comparison against probing a standard model, which is the cheapest competing explanation.
Where This Sits
Latent bottleneck fusion supplies the method. The Multimodal Bottleneck Transformer (Attention Bottlenecks for Multimodal Fusion) makes video and audio streams exchange information only through $B$ shared tokens. This cuts cross-stream cost from $O(N_A N_B)$ to $O(B(N_A+N_B))$. The same paper found that fusing at a middle layer beats early fusion. The Denoising Bottleneck Fusion model (Denoising Bottleneck with Mutual Information Maximization for Video Multimodal Fusion) argues that such a channel filters frame noise, and adds a contrastive term so that useful detail is not lost. Neither paper measures uncertainty, and neither touches text retrieval. We take their architecture and ask the question they skip.
Three papers constrain the design. The calibration-bottleneck study (Calibration Bottleneck) shows, on ResNets, that heavily compressed top layers are hard to fix with temperature scaling. On CIFAR-100, weight decay worsened calibrated ECE from 0.83 to 1.03, while its layer-freezing remedy reached 0.58. Here ECE means expected calibration error, the average gap between stated confidence and actual accuracy. That result predicts our bottleneck can fail in precisely the way we care about, so $B$ must be swept rather than fixed. The distractor study (Mind the Confidence Gap: Overconfidence, Calibration, and Distractor Effects in Large Language Models) shows that the direction of the effect depends on difficulty. Structured distractors improve calibration on hard questions but worsen it on easy TriviaQA questions, so any evaluation has to stratify by difficulty. The retrieval-triggering study (When Do LLMs Need Retrieval Augmentation?) finds that overconfident models are the ones most easily distracted. It also finds that some models, such as Vicuna-v1.5-7B, do not respond to prompt-based fixes at all, which motivates a change inside the model.
The closest prior work is two retrieval papers, and both operate outside the model. IB-RAG (An Information Bottleneck Perspective for Effective Noise Filtering on Retrieval-Augmented Generation) trains a separate filter model. It compresses passages to about 2.5% of their length under an information-bottleneck objective, a loss that trades compression against predictive value. It never reports calibration. CalibRAG (Reliable Decision‑Making via Calibration‑Oriented Retrieval‑Augmented Generation) reranks documents by the predicted probability that a decision grounded in them is correct, then passes the chosen document to the generator unchanged. Our proposal differs from both in two ways. The compression happens in the generator's own attention, and the confidence estimate comes from the compressed channel itself rather than from a separate model.
Research Program
The base model is Llama-3.1-8B-Instruct, the backbone used in the Ext2Gen noise experiments (Aligning Extraction and Generation for Robust Retrieval-Augmented Generation), adapted with LoRA, a method that trains small low-rank weight updates. The bottleneck is built entirely from the attention mask plus $B$ learned embeddings.
Below the fusion layer $L_f$, passage tokens and question tokens cannot see each other. From $L_f$ upward, passage tokens attend only to their own passage and the latents. Question and answer tokens attend only to themselves and the latents. The latents attend to everything, including the question, so the compression is conditioned on what is being asked. That conditioning mirrors IB-RAG's conditional objective, applied inside the network instead of on text strings.
A small head pools the final-layer latent states into a correctness probability $c$. The answer loss is standard next-token prediction on the gold answer given noisy context. The head is trained with log loss $-[y\log c + (1-y)\log(1-c)]$, where $y$ indicates whether a sampled answer was judged correct. We also keep a divergence penalty toward the frozen model's output distribution to limit capability drift. This follows the regularization used in Large Language Models Must Be Taught to Know What They Don't Know. That paper reduced open-ended MMLU ECE from about 30% to 10% with roughly a thousand graded examples.
Training contexts mix gold passages with between 0 and 25 irrelevant retrieved chunks, following the Ext2Gen protocol. They also include answer-absent contexts, where no passage contains the answer.
The analysis has three parts. First, we sweep $B$ and $L_f$ and trace accuracy, raw ECE, and calibrated ECE after temperature scaling. The last of these is the calibratability measure from the vision study, so we can test whether its U-shape reappears along the width axis. Second, we measure confidence drift. For a fixed question, we raise the number of distractors and compare how much confidence falls with how much accuracy falls. A calibrated system should move the two together. Third, we test whether the latent-token confidence distinguishes two situations: "the context lacks the answer" and "the model lacks the knowledge". We do this by comparing answer-absent contexts on questions the closed-book model answers correctly against the same contexts on questions it misses.
Baselines
The naive reference point is the unmodified instruct model reading the top-k retrieved passages, with confidence taken as the sequence probability of its answer. The same model with no retrieval gives the closed-book floor. On NQ, IB-RAG reports that plain top-5 retrieval reaches 14.7 exact match (EM, the fraction of answers matching the gold string exactly), below the 16.2 of no retrieval. We expect to reproduce that ordering under our noise mixtures.
The strongest competitor on calibration is a correctness probe, a small classifier trained on a standard model's internal activations to predict whether its answer is right. Internal activations are known to carry error signals even when the output probabilities are overconfident (Hao Liu et al., cited in the file only by author name and not linked here). We pair the probe with LoRA uncertainty fine-tuning in the style of Large Language Models Must Be Taught to Know What They Don't Know, trained on exactly our noisy contexts. This is the strongest threat because it has the same white-box access, the same data, and the same kind of calibration objective, and differs only in the missing bottleneck. It will receive the same LoRA rank, the same graded examples, the same number of tuning trials, and the same temperature-scaling step.
The strongest competitor on accuracy under noise is an extract-then-generate pipeline. We use the IB-RAG filter feeding the same generator, with a probe and temperature scaling added so that it produces a confidence at all. IB-RAG reports NQ EM of 21.5 and a True-Flip-Rate of 17.6%, down from 51.0% with standard filtering. The True-Flip-Rate counts correct closed-book answers that the retrieved context flips to wrong. Ext2Gen reports an average score of 0.557 against 0.473 for the default Llama-3.1-8B model under heavy noise. On HotpotQA, CalibRAG reports 42.37% accuracy against 34.98% for cross-encoder reranking. We run CalibRAG-style reranking as the retrieval-side calibration baseline.
We found no reported ECE or Brier score on Lacuna for IB-RAG, Ext2Gen, or CalibRAG on NQ or HotpotQA. The calibration comparison therefore relies on our own reruns under matched settings rather than on published calibration figures.
The ablation that isolates the key ingredient keeps the same LoRA budget and the same correctness head but removes the mask, so passages attend to the question stream in full. It also reads confidence from the final answer-token state instead of from the latents. A second ablation keeps the bottleneck but reads confidence from the answer tokens. This separates "the bottleneck changes the representation" from "the bottleneck is the right place to read uncertainty".
To count as beating the probe baseline, the proposal must show lower ECE and Brier score, and higher correctness AUROC (how well confidence ranks right answers above wrong ones), in the 10- and 25-distractor conditions and in answer-absent contexts. The difference must exceed variation across seeds. On accuracy, it must stay within a small margin of the extract-then-generate pipeline.
Why not just…?
Why not just prompt the model to abstain or to doubt itself? Punish+Explain prompting raised LLaMA-2's rate of declaring itself uncertain from 16.8% to 69.2% (When Do LLMs Need Retrieval Augmentation?). DAS negation instructions contributed to large EM gains (Improving Zero-shot Reader). However, the same studies show that some models do not respond to such prompts. Separately, verbalized confidence shifts with prompt wording while accuracy stays the same (Large Language Models Must Be Taught to Know What They Don't Know). We include both prompts as baselines. If they match our calibration numbers under 25 distractors, the architecture is not justified.
Why not just filter the text first and temperature-scale the output? This is a strong accuracy competitor, and IB-RAG's flip-rate reduction is large. Its weaknesses are an extra model call and, as Ext2Gen reports, the fact that separate extractors underperform when trained apart from the generator. Most important for our question, a filtered snippet of about twelve words may strip out exactly the hedging cues that confidence needs. That is the over-compression failure predicted by the calibration-bottleneck study. The matched pipeline baseline settles this directly. If filtering plus a probe plus temperature scaling matches our calibration, our remaining advantage would be single-pass compute only, and we would report it that way.
Why not just probe or calibration-fine-tune the standard model on noisy data? This is the real threat. It is also untested in the literature, which reports these methods only on clean inputs. If the probe on a full-attention model degrades no faster with distractors than the latent-token head, then the bottleneck adds nothing to uncertainty estimation. The no-mask ablation measures exactly this.
Decisive Checks
We evaluate on NQ, HotpotQA, and TriviaQA with controlled distractor counts (0, 5, 10, 25) plus answer-absent variants, on SQuAD 2.0 for unanswerable questions, and on UDA FinHybrid as a hard transfer set. On FinHybrid, GPT-4 falls from 45.9 with retrieval to 0.4 when reading the full document (UDA). All results are stratified by closed-book difficulty so the TriviaQA reversal can be detected.
Three results would decide the proposal. First, the gap in calibrated ECE between latent-token confidence and the full-attention probe should widen as distractors increase. If the gap is flat, the claim fails. Second, calibrated ECE across $B \in \{4, 16, 64, 256\}$ should trace a U-shape with an interior minimum. A monotone curve would mean the calibratability concern from vision does not carry over, or dominates entirely, and either outcome is informative. Third, on answer-absent contexts, latent-token confidence should fall even for questions the model knows closed-book, while the answer stays correct. That pattern would show the channel measures evidence from the context rather than overall correctness, which is the property no output-level method provides.
References
- Improving Zero-shot Reader by Reducing Distractions from Irrelevant Documents in Open-Domain Question Answering
- Calibration Bottleneck: Over-compressed Representations are Less Calibratable
- Attention Bottlenecks for Multimodal Fusion
- Denoising Bottleneck with Mutual Information Maximization for Video Multimodal Fusion
- Mind the Confidence Gap: Overconfidence, Calibration, and Distractor Effects in Large Language Models
- When Do LLMs Need Retrieval Augmentation? Mitigating LLMs' Overconfidence Helps Retrieval Augmentation
- An Information Bottleneck Perspective for Effective Noise Filtering on Retrieval-Augmented Generation
- Reliable Decision‑Making via Calibration‑Oriented Retrieval‑Augmented Generation
- Aligning Extraction and Generation for Robust Retrieval-Augmented Generation
- Large Language Models Must Be Taught to Know What They Don’t Know
- UDA: A Benchmark Suite for Retrieval Augmented Generation in Real-World Document Analysis