The Long-Context Reader as a Manipulable Mechanism: Median Aggregation Against Strategic Distraction

When a language model answers a question from many retrieved passages, it reads them all at once. Its attention is a soft weighting over every token, and no passage is ever fully switched off. Each passage therefore pulls on the answer, and a passage that is longer, more repetitive, or cleverly worded pulls harder. This is the same flaw that makes an average a poor way to pick a meeting point: whoever reports the most extreme position drags the result toward themselves. Mechanism design without money has a standard fix for the averaging problem. Pick the median, and no participant can move the outcome toward themselves by exaggerating.

This note proposes treating the reader as that kind of mechanism. For questions whose answers are numbers (dates, counts, amounts, measured quantities), each source is read in isolation, a single number is extracted from it, and the numbers are combined with a median rule that includes "phantom" default points. Under this rule, distraction planted inside one source can move only that source's own point. A source that wants the answer to land near its preferred value gains nothing by padding or adding distractors. The empirical question is whether this bounded-influence property survives the reality of noisy LLM extraction, and what it costs in accuracy on honest inputs.

Core Claim

The claim has two parts. The first is a measurement. For a full-context reader, the influence a single source has on the output grows with how much distracting text that source contains. We will call this the manipulation gain. The second is a design. For numeric-answer questions, an isolate-extract-median reader bounds each source's influence to one rank position, regardless of that source's content. It also gives each source no incentive to misrepresent its preferred value. This comes at a measurable but modest cost in clean accuracy on questions that do not need cross-document reasoning.

The novelty is not the median itself, which is a textbook robust estimator. It is two things together. We recognise that the context-dilution failure documented in the distraction literature is exactly a manipulability failure in the mechanism-design sense. We then quantify it with an influence metric that the robustness benchmarks do not report.

Grounding in Prior Work

Four papers motivate the proposal by establishing that distraction is severe and not fixable by instruction.

All four treat noise as passive. None asks how much one self-interested source can steer the output.

Two lines of mechanism-design work supply the method. The Capacity Constrained Facility Location Problem characterises generalized median mechanisms as the only anonymous strategyproof rules in their capacitated one-dimensional setting. These rules add $n+1$ fixed phantom points to $n$ reports and take the median. We reuse this rule family directly, with the answer line as the facility location. Random Rank supplies the randomized counterpart that we test as a secondary rule for non-numeric answers. A Dictatorship Theorem for Cake Cutting and Incentive Compatible Two Player Cake Cutting constrain the design. They show that truthfulness without money carries an intrinsic welfare price, bounded near 93% of optimum in the cake-cutting case. We should therefore expect, and report, a clean-accuracy cost rather than claim a free lunch.

The closest existing bridge is Error-Aware Reverse Auction Mechanism for Large Language Model Routing, known as EA-RAM. It is a paid, incentive-compatible auction for choosing which model answers a query. It models noise in providers' self-predicted success and in the external evaluator's scoring, but it assumes a fixed query broadcast to all providers. We differ on every axis that matters:

  • What is allocated: EA-RAM allocates the job; we allocate influence over the answer.
  • Who is strategic: in EA-RAM it is model providers; here it is content sources.
  • Money: EA-RAM uses payments; we use none.
  • Where the noise sits: EA-RAM treats the prompt as fixed and models noise elsewhere; we locate it inside the prompt.

Research Program

Setting. A question $q$ has a numeric answer. There are $n$ sources, and source $i$ holds a preferred value $t_i$: its honest belief, or the value it wants the system to output. A source is harmed in proportion to $|y - t_i|$, where $y$ is the system's output. This is the one-dimensional, single-peaked setting in which median rules are strategyproof.

The full-context reader. The standard reader outputs $y = R(q, s_1, \dots, s_n)$ from the concatenated passages $s_i$. We define the manipulation gain of source $i$ as the largest shift in $y$ toward a target that source $i$ can achieve by rewriting only $s_i$ within a token budget. We estimate it with the error-guided tree search from Breaking Focus, redirected from "make the model wrong" to "make the model output value $v$."

The proposed reader. It computes $x_i = E(q, s_i)$ separately for each source. Here $E$ is the same LLM prompted with the DAS-style "return unanswerable" escape. Abstentions are dropped, leaving $m$ answering sources. The output is

$$y = \mathrm{median}(x_1, \dots, x_m, p_1, \dots, p_{m+1}),$$

where the phantoms $p_j$ are fixed before any source is read. In our main variant they are split between the model's closed-book answer and the boundaries of a plausible range. That split sets how much the system trusts its prior when few sources answer. Because the phantoms and the rule never depend on source content, a rewritten $s_i$ can change only $x_i$.

Theoretical claims. Two statements follow from standard order-statistic arguments. First, if the extractor is exact, reporting $t_i$ is a dominant strategy. Second, $f$ manipulating sources can push $y$ no further than the range spanned by the honest points and phantoms, as long as they are outnumbered.

Extending to noisy extraction. The less obvious analysis concerns extractor error, which is the regime Breaking Focus shows is real even within a single passage. Suppose each $x_i$ lies within $\epsilon_i$ of what source $i$ genuinely asserts. Then the output is within $\max_i \epsilon_i$ of the ideal median. We will test whether $\epsilon_i$ itself grows when a source pads its own passage. If it does, that is the one channel through which content still buys influence. Measuring its size is part of the contribution.

Non-numeric answers. For multiple-choice questions we test Random Rank's categorical analogue: a uniformly random answering source is chosen as the dictator. It is strategyproof but high-variance, and we compare it against plurality voting.

Baselines

Naive full-context reader. This is the do-nothing reference: the same LLM given all passages concatenated. It anchors both accuracy and manipulation gain. It is run with the identical model, decoding settings, and passage sets as every other method.

DAS. This is the strongest training-free distraction defence in the file. Improving Zero-shot Reader reports an average 64% Exact Match increase over naive zero-shot readers, and a 231% increase on SQuAD with 100 retriever documents. It matters specifically because its re-scoring weights answers by how likely the passage makes the query. A strategic source can inflate that score by echoing the query, so DAS may be accurate yet manipulable. We run it with its published prompt and scoring on our model. Beating it means lower manipulation gain at equal or better clean accuracy.

CompAct. This is the strongest compression defence. CompAct reports a 7.0 F1 improvement over the RECOMP compressor on HotpotQA. We give it the same passages and the same total LLM-call budget as the proposal.

DPO-trained robust reader. This follows Breaking Focus: preference training on distracted-versus-clean pairs. It is the strongest learned competitor. No reported number was found on Lacuna for its restored accuracy. We train it on distractor pairs drawn from a held-out split of the same generator and evaluate it against an adaptive attacker that re-searches against the trained model.

Ablations of the key ingredient. Isolated extraction followed by an unweighted mean, and isolated extraction followed by DAS-score weighting, both keep the per-source isolation but replace the median. This separates the gain from isolation (distraction cannot cross passages) from the gain from the aggregation rule (a single extreme report cannot drag the result). A phantom-free median isolates the contribution of phantoms under abstention-heavy inputs.

All methods share the base model, a matched tuning budget on a common validation split, and a reported count of LLM calls, since isolation costs $n$ calls.

Why not just…?

Why not just prompt the reader to ignore irrelevant or promotional content? Breaking Focus reports that such prompting is ineffective against coherent distractors. Our attacker is adaptive, so an instruction is simply one more thing the search can route around. We include "ignore irrelevant information" as a prompt variant of the naive reader to confirm this in the manipulation-gain setting.

Why not just compress or rerank before reading? This is the most serious threat. Against non-adaptive noise, CompAct or a reranker may match our accuracy. However, the compressor is itself an LLM reading all passages jointly. A source that controls its text can target the compressor as easily as the reader, and a reranker's relevance score is exactly what a source optimises for. The experiment settles it by running the adaptive attacker against the full compressor-plus-reader pipeline. If CompAct's manipulation gain stays as low as the median reader's, the incentive framing adds little and we will say so.

Why not just call this robust statistics and use self-consistency or majority voting? Partly we should. The median is a robust estimator, and per-document voting exists. What the mechanism-design lens adds is threefold:

  • It frames the goal as removing each honest source's reason to pad, not only resisting adversaries.
  • It supplies the phantom family for handling abstentions and priors.
  • It gives the negative results that bound what any non-monetary rule can achieve.

Plurality voting over isolated extractions is in our baselines. If it matches the median on numeric answers, the contribution reduces to the influence measurement.

Why not just pay sources, as EA-RAM does? Web pages and departmental documents cannot be paid or penalised. EA-RAM itself notes that penalising bad outputs needs pre-funded reserves. Its incentive guarantee also assumes that participants' self-predictions are reasonably calibrated, which is the kind of guarantee that the calibration collapse in Breaking Focus puts in doubt.

Proposed Experiments

Data. We build a strategic-source benchmark from numeric-answer questions. These come from UDA's document-analysis tasks, from GSM8k within the six RouterBench sub-benchmarks used by EA-RAM, from MATH-500, and from the numeric and date subsets of Natural Questions and TriviaQA with retrieved passages, following the DAS setup.

Sources. Each question receives $n \in \{5, 10, 20\}$ sources, mixing gold-bearing passages, neutral retrieved passages, and $f$ strategic passages. The strategic passages are rewritten by the targeted tree search to push toward a random target value under token budgets from 1× to 8× a normal passage length.

Metrics. We report four:

  • exact-match and relative-error accuracy on clean inputs;
  • accuracy under attack;
  • manipulation gain, measured as the median shift toward the target per manipulating source;
  • a padding-response curve, meaning how manipulation gain varies with the strategic source's token budget.

Decisive checks. Four results would settle the proposal:

  1. The naive and DAS readers' manipulation gain rises with padding budget, while the median reader's stays flat. This is the central prediction.
  2. The mean ablation is manipulable while the median is not. This shows the rule, not isolation alone, does the work.
  3. The median reader's clean-accuracy cost stays small on single-source numeric questions. We will measure it separately on HotpotQA's distractor setting, where combining two documents is required, expecting isolation to fail there. That failure defines the method's scope.
  4. Within-passage padding inflates the extractor error $\epsilon_i$. This would quantify the residual channel that no aggregation rule can close.

Categorical transfer test. We run the Random Rank dictator and plurality voting on the Breaking Focus multiple-choice suite (MMLU, CommonsenseQA, OpenbookQA) to see whether bounded influence carries over beyond numeric answers.