{"id":"65347aa7-b0ef-417c-93be-d81559c79018","arxiv_id":"2607.11783","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":5.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"RAG transfers ideological discourses from retrieved COVID-19 treatment texts into LLM answers, with strongest alignment at moderate sampling temperatures and weaker transfer when sampling is overly deterministic.","lead":"This study finds that retrieval-augmented generation can pass ideological language from its knowledge base into LLM answers, and that sampling temperature changes how strongly that happens. Moderate temperatures transfer discourse most; very low temperatures suppress it. The work matters for anyone using RAG on contested topics.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"Similarity scores to LMDA references drawn from the same RAG corpus may capture ordinary retrieval fidelity rather than ideology-specific discourse transfer, leaving the temperature effect unisolated.","rationale":"The reader correctly identifies the LMDA isolation of three discourses as the weakest assumption. The concern raised here is adjacent: even granting that the dimensions are stable, the experimental design (shared corpus for KB and references, exclusively ideological questions) fails to isolate ideology transfer from ordinary RAG content fidelity. Because only the abstract is available, neither issue can be resolved; the CONDITIONAL verdict with low confidence therefore remains appropriate. No stronger objection is justified on the given evidence, and the temperature finding would still be of interest if subsequent controls confirm specificity.","tokens_in":2025,"tokens_out":437,"duration_ms":16983,"concrete_test":"Using the same RAG pipeline and temperature grid, generate answers to a matched set of non-ideological factual questions drawn from the identical corpus; recompute the identical semantic and lexical similarities to the three LMDA reference texts. If the moderate-temperature peak and low-temperature drop reappear at comparable magnitude, the pattern is not ideology-specific and the headline claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that RAG transfers ideological discourses, with alignment peaking at moderate temperatures and dropping at low temperatures—depends on semantic and lexical similarity of generated answers to three LMDA-derived reference texts. Those references come from the identical 1,117-article COVID-19 corpus that serves as the RAG knowledge base. Successful retrieval therefore predicts elevated similarity by construction; designating the LMDA dimensions “ideological discourses” does not automatically convert that similarity into a pure measure of ideology transfer. The abstract reports no controls that would separate ideology from topical, lexical, or stylistic overlap (e.g., neutral factual questions, non-ideological KBs, or discourse-scrambled references). Consequently the observed temperature modulation cannot yet be attributed specifically to ideological discourse rather than to the generic interaction of sampling temperature with retrieval-augmented generation.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The manuscript argues that Retrieval-Augmented Generation (RAG) can transfer ideological discourses present in the retrieved corpus into LLM outputs, and that sampling temperature measurably modulates this transfer. Using Lexical Multidimensional Analysis (LMDA) on a corpus of 1,117 COVID-19 treatment articles, the authors identify three ideological discourses; the same corpus then serves as the RAG knowledge base. Several LLMs answer ideological questions across sampling temperatures; generated answers are scored for semantic and lexical similarity to the LMDA-derived reference texts. The abstract reports that discursive alignment peaks at moderate temperatures (where stochasticity and retrieval grounding are balanced) and drops at low temperatures (where deterministic sampling suppresses discourse transfer).","tokens_in":2163,"tokens_out":842,"duration_ms":13632,"significance":"If the result holds under proper controls, the work would be a useful contribution to the study of bias propagation in RAG systems: it links a concrete generation hyperparameter (temperature) to the strength of ideological discourse transfer and supplies a corpus-driven pipeline (LMDA + similarity metrics) that others could reuse. The temperature finding is potentially actionable for practitioners who wish to dampen or surface retrieved ideological content. Because only the abstract is available, however, the magnitude, robustness, and generality of the effect cannot yet be assessed; significance therefore remains provisional pending full methods, results, and controls.","major_comments":[{"comment":"Abstract (pipeline description): The three LMDA reference texts are extracted from the identical 1,117-article COVID-19 corpus that constitutes the RAG knowledge base. Elevated semantic/lexical similarity of generated answers to those references is therefore expected under ordinary successful retrieval; designating the LMDA dimensions “ideological discourses” does not by itself convert that similarity into a pure measure of ideology transfer. The abstract mentions no controls that would isolate ideology from topical, lexical, or stylistic overlap (e.g., neutral factual questions, non-ideological KBs, discourse-scrambled or shuffled references, or a non-RAG baseline). Without such isolation the central claim—that RAG transfers ideological discourses and that temperature modulates that transfer—remains under-supported.","section":"Abstract"},{"comment":"Abstract (results claim): The abstract asserts that “discursive alignment … is highest at moderate temperatures and drops at low temperatures,” yet supplies no temperature grid, model list, quantitative similarity scores, error bars, statistical tests, or effect sizes. From the available text it is impossible to judge whether the reported temperature dependence is robust, statistically reliable, or large enough to matter. This is load-bearing for the paper’s second main claim.","section":"Abstract"}],"minor_comments":[{"comment":"Abstract: The term “discoursive alignment” appears once; elsewhere “discursive” is used. Standardize spelling.","section":"Abstract"},{"comment":"Abstract: “several LLMs” and “different sampling temperatures” are left unspecified; even a brief parenthetical list of models and the temperature range would improve readability of the abstract itself.","section":"Abstract"},{"comment":"Abstract: The phrase “the RAG framework, comprising ideological discourses” is slightly awkward; clarifying that the discourses reside in the retrieved documents rather than in the RAG architecture would reduce ambiguity.","section":"Abstract"}],"recommendation":"uncertain","confidential_remarks":"Only the abstract was supplied for review; the full manuscript (methods, tables, figures, statistical tests, model and temperature details, and any controls) was unavailable. The recommendation is therefore “uncertain.” If the full paper contains the missing controls and quantitative results, the methodological concern may be resolvable and the paper could move to minor or major revision; if those elements are absent, the circularity risk identified by the stress-test note is load-bearing and would justify major revision or rejection. I recommend the editor obtain the full text before a final decision."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing to know: this abstract reports a clean empirical interaction—RAG can pull LMDA-identified ideological discourses from a COVID-treatment corpus into answers, and that alignment peaks at moderate sampling temperatures and drops when sampling is too deterministic. That is a useful angle for RAG evaluation and product defaults on contested topics. It is not a pure rehash; prior work treated RAG robustness and LLM political bias separately, so the temperature sweep on discourse transfer is a legitimate new measurement even if the broader program is familiar.\n\nWhat they do well on paper is the pipeline sketch: LMDA on 1,117 articles → three discourses as reference texts → RAG answers across temperatures → semantic/lexical similarity. The claim is coherent and the free parameters (temperature grid, three dimensions) are explicit. No formal proofs or released artifacts here, but the design is not definitionally circular—discourses are extracted first, then generation is scored against references.\n\nThe soft spot is real and load-bearing, not manufactured. The stress-test lands: those reference texts come from the same corpus that is the RAG knowledge base. Elevated similarity is what successful retrieval predicts by construction. Calling the LMDA dimensions “ideological discourses” does not by itself prove the temperature effect is ideology-specific rather than generic retrieval-plus-sampling behavior. The abstract mentions no controls (neutral questions, non-ideological KBs, scrambled references, or baselines that hold topic fixed while removing discourse). Without those, and without numbers, error bars, model list, or tests, we cannot yet attribute the peak to ideology transfer. Circularity risk is moderate, not fatal; it is just unisolated.\n\nThis is for people who build or evaluate RAG systems on contested domains and for AI-safety folks tracking bias transmission. A serious referee should see the full paper if the quantitative results, controls, and data/code exist; the finding would matter for evaluation design even if the ideology framing needs tightening. I would not cite from the abstract alone. Send it to peer review rather than desk-reject—conditional acceptance is the right posture until the methods and numbers are on the table.","headline":"Abstract-only claim that RAG transfers LMDA-derived COVID discourses with a moderate-temperature peak is plausible and field-relevant, but the design may just be measuring ordinary retrieval fidelity.","tokens_in":2817,"tokens_out":525,"would_cite":false,"duration_ms":7774,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Retrieval-augmented generation transfers ideological discourse into LLM answers, with transfer strength peaking at moderate sampling temperatures.","keywords":["retrieval-augmented generation","ideological bias","sampling temperature","large language models","discourse transfer","Lexical Multidimensional Analysis","COVID-19","LLM generation"],"falsifier":"Replicate the pipeline on a different corpus with clear ideological poles (for example left- and right-leaning political news), extract dimensions the same way, and test whether moderate temperatures still produce the highest lexical and semantic alignment with those reference discourses; if alignment is flat or highest at low temperature, the temperature-transfer claim fails.","tokens_in":2896,"feed_emoji":"🌡️","tokens_out":784,"duration_ms":16100,"temperature":0.7,"pith_summary":"This paper sets out to show that retrieval-augmented generation is not ideology-neutral: when the external knowledge source contains ideological positions, those discourses are transferred into the LLM’s answers. Using a corpus of 1,117 COVID-19 treatment articles, the authors extract three ideological discourses via Lexical Multidimensional Analysis and then measure how closely generated answers align with those reference texts at different sampling temperatures. Alignment is highest at moderate temperatures, where the model balances stochasticity with retrieval grounding, and falls at low temperatures, where deterministic sampling suppresses discourse transfer. A reader who accepts the result would care because RAG is widely used precisely to ground models in external knowledge, yet that same grounding can quietly inject ideology rather than pure fact. Temperature therefore becomes a practical lever that either amplifies or mutes the ideological coloration of RAG outputs.","feed_headline":"Moderate temperature maximizes ideological transfer in RAG","feed_subtitle":"Low sampling temperatures suppress discourse transfer from retrieved texts into LLM answers.","key_machinery":"Lexical Multidimensional Analysis applied to 1,117 COVID-19 treatment articles, isolating three ideological discourses that serve as reference texts; RAG then answers ideological questions at varied sampling temperatures, with semantic and lexical similarity scores quantifying discourse transfer.","core_discovery":"The RAG framework is prone to transferring ideological discourses from retrieved material into LLM responses, and sampling temperature has a measurable impact: discursive alignment with ideological reference texts is highest at moderate temperatures and drops at low temperatures, where overly deterministic sampling suppresses discourse transfer.","pith_inferences":["Similar transfer effects may appear for other latent dimensions beyond ideology, such as commercial framing or partisan stance in news corpora.","Temperature schedules or multi-temperature ensembles could be designed to attenuate discourse transfer while still preserving factual grounding.","The three discourses identified in the COVID-19 treatment literature may themselves prove corpus-specific rather than universal, inviting replication on other domains.","If the pattern holds across domains, retrieval-corpus curation becomes as important as temperature for controlling output ideology."],"forward_implications":["RAG systems built on ideologically mixed corpora will systematically color LLM answers with those discourses rather than remaining neutral.","Very low temperature settings reduce ideological transfer but may also reduce the practical benefit of retrieval grounding.","Moderate temperature settings maximize the injection of retrieved ideological framing into generated answers.","Temperature can serve as a production control knob for the strength of discourse transfer in deployed RAG systems.","Evaluations of RAG factuality that ignore ideological alignment will miss a systematic bias channel."],"fun_headline_variants":["Moderate temps peak ideological transfer from RAG sources","Low sampling heat suppresses discourse transfer in RAG","RAG leaks ideologies strongest at mid-range temperatures","Temperature gauges how RAG injects ideology into answers","Deterministic sampling mutes ideological transfer via RAG"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"That the three dimensions extracted by Lexical Multidimensional Analysis from the COVID-19 treatment articles are genuine, stable ideological discourses that validly serve as reference texts for measuring transfer.","fun_headline_variants_meta":{"raw":{"variants":["Moderate temps peak ideological transfer from RAG sources","Low sampling heat suppresses discourse transfer in RAG","RAG leaks ideologies strongest at mid-range temperatures","Temperature gauges how RAG injects ideology into answers","Deterministic sampling mutes ideological transfer via RAG"]},"model":"grok-4.5","effort":"low","cost_usd":0.00285,"raw_usage":{"total_tokens":1035,"prompt_tokens":744,"num_sources_used":0,"completion_tokens":71,"cost_in_usd_ticks":28500000,"prompt_tokens_details":{"text_tokens":744,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":220,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":744,"tokens_out":71,"duration_ms":2511,"temperature":1.0,"reasoning_tokens":220,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T03:12:43.278950+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Replicate the pipeline on a different corpus with clear ideological poles (for example left- and right-leaning political news), extract dimensions the same way, and test whether moderate temperatures still produce the highest lexical and semantic alignment with those reference discourses; if alignment is flat or highest at low temperature, the temperature-transfer claim fails.","supporting_citations":[],"review_version":1}