{"id":"eb30e641-8a3c-4432-b9d6-7f68fd48992c","arxiv_id":"2607.17766","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Selectively passing evidence-grounded paper terminology into a streaming speech translator improves technical-term recall and translation quality more than injecting whole-document context.","lead":"When translating a technical talk live, this paper feeds the translator only the paper-specific terms (model names, acronyms, datasets) that the current snippet actually mentions, instead of injecting the whole document. On two technical-talk benchmarks the selective method beats always-on context on terminology recall and overall quality, at equal or slightly lower latency.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Primary terminology metrics use gold terms from the same extractor that builds the memory; external GoldTag on En-De actually drops, undercutting the claimed consistent gains.","rationale":"The central claim is that context should be converted into sparse, evidence-grounded terminology adaptation, with consistent gains on both language directions and on an external benchmark. This rests on the terminology metrics. If the primary metric's gold set is the same extractor output that drives adaptation, the headline gains are partly a measure of self-consistency, not external validity. The paper's own conservative diagnostic shows the effect is smaller, and the one truly external official metric (GoldTag) shows a decrease on En-De, contradicting the 'consistent gains' language. This does not necessarily invalidate the framework—EGTA still beats uniform context and improves many metrics—but it means the strongest claims are not supported as stated. The reader's weakest assumption is closely aligned, so the conditional verdict stands; the required revision is to report reference-derived metrics as primary and disclose the negative GoldTag cell.","tokens_in":14257,"tokens_out":6679,"duration_ms":70809,"concrete_test":"Recompute Table 2 and Table 9 with a gold set derived from the official references (e.g., all paper terms that appear in the reference, or the official GoldTag) instead of the Qwen-extracted memory. Specifically, compare EGTA-RG B=2 against baseline on ACL60/60-dev En-De using the official GoldTag matching script; if the result reproduces the 0.656→0.642 decrease, the claim of consistent external gains is false for that direction and the abstract/main text must be revised. Also report the reference-supported micro-F1 from §A.8 as the primary terminology metric for MCIF-dev.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.3 defines NERclean recall 'over cleaned paper-specific entities' but the gold set is not derived from references; Section 3.2 constructs the terminology memory using a prompt-based Qwen3-30B-Instruct extractor, and Tables 6-7 show the metric categories are the memory's own categories. Since EGTA-G applies logit bias only to those memory terms (Eq. 2) and EGTA-R exposes them in the prompt, a large part of the reported +0.212/+0.266 NERclean gain (Table 2) may reflect the model echoing the extractor's vocabulary rather than genuine terminology fidelity. The paper's own 'reference-supported terminology matching' (§A.8) shows a smaller effect: micro-F1 from 0.625 to 0.839 on En-Zh and 0.575 to 0.753 on En-De. More directly, the external official GoldTag recall on ACL60/60-dev (Table 9) decreases on En-De from 0.656 to 0.642, contradicting the abstract's 'consistent terminology-recall gains.' Thus the direction and magnitude of the headline terminology gains are not robust to the choice of gold set; the central claim of consistent selective-adaptation gains is not firmly established without an independent, reference-derived primary metric.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that in simultaneous speech translation of technical talks, extra document context is most useful when converted into a sparse, evidence-conditioned terminology-adaptation signal rather than injected uniformly. It proposes EGTA, which builds a document terminology memory with a prompt-based LLM extractor, selects compact candidate terms per streaming segment using surface-form matching, and applies them through ASR/speech-side conditioning (R) and decoder-side logit bias (G). On MCIF-dev and ACL60/60-dev, EGTA-RG is reported to improve BLEU, XCOMET-XL, NERclean recall, acronym recall, and latency over a shared end-to-end backbone, with shuffled-memory controls and activation audits intended to establish evidence grounding.","tokens_in":14389,"tokens_out":2144,"duration_ms":25283,"significance":"If the central claim holds, the paper makes a useful practical contribution: an inference-time, architecture-agnostic method for terminology adaptation in streaming technical-talk translation that does not require fine-tuning. The experimental design has strengths: a shared backbone across variants, matched latency operating points, uniform-context and Global-G stress baselines, a shuffled-memory control, and paired bootstrap at segment and talk level with p<0.001 for the primary terminology metrics on MCIF-dev. The activation audits and evidence-strength bucket analysis are also promising. However, the primary terminology metrics are computed against gold terms produced by the same extractor that constructs the adaptation memory, making part of the headline gain a self-consistency effect. The external official GoldTag recall on ACL60/60-dev En→De decreases, contradicting the abstract's claim of 'consistent terminology-recall gains.' The core lesson — that selective evidence-conditioned terminology adaptation is superior to uniform context — is plausible and interesting, but its empirical support is not yet robust without an independent, reference-derived primary metric.","major_comments":[{"comment":"NERclean recall and acronym recall are computed over 'cleaned paper-specific entities' whose gold categories come from the same Qwen3-30B-Instruct extractor that builds the terminology memory TD (Tables 6–7 clearly show the metric categories are the memory's own categories). Since EGTA-G biases only tokens in V(T_i) and EGTA-R exposes those terms in the prompt, a large part of the +0.212/+0.266 NERclean gain (Table 2) may simply reflect the model echoing the extractor's vocabulary. The paper never states that the NER gold is independently annotated; §A.8 introduces reference-supported matching as a separate 'conservative' diagnostic, implying NERclean's gold is not reference-grounded. This circularity undermines the primary terminology-fidelity claims.","section":"§4.3, §3.2, Tables 6–7"},{"comment":"The abstract and §5.1 claim 'consistent terminology-recall gains' on ACL60/60-dev without fine-tuning, but Table 9 shows that on the official GoldTag terminology recall, En→De decreases from 0.656 (baseline) to 0.642 (EGTA-RG B=2), while uniform context gives 0.631. The paper does not acknowledge this drop or explain why the self-constructed NERclean/acronym metrics improve while the official tagged-terminology metric does not. Since the central claim is consistency of terminology gains, a single external metric moving in the opposite direction is load-bearing and must be addressed.","section":"Table 9 and Abstract"},{"comment":"The reference-supported terminology matching diagnostic reports micro-F1 gains from 0.625→0.839 (En→Zh) and 0.575→0.753 (En→De), which are smaller in relative terms than the NERclean gains and are explicitly described as 'conservative.' This confirms that the magnitude of the terminology benefit is sensitive to the choice of gold set. The paper should make the reference-supported metric a primary result rather than a complementary appendix item, or provide a direct comparison of all terminology metrics on the same gold.","section":"§A.8"},{"comment":"The statistical robustness section (Table 8) reports p<0.001 only for NERclean and acronym recall on MCIF-dev. For ACL60/60-dev, significance is reported for En→Zh XCOMET-XL and acronym/tagged recall, but no significance is reported for the En→De GoldTag drop. The claim of 'consistent' gains requires either reporting the GoldTag significance or correcting the claim. Without this, the external-validation argument is incomplete.","section":"§5.1/A.5"}],"minor_comments":[{"comment":"The reference list contains 'V oita' instead of 'Voita' (misspacing in the rendered text).","section":"§1/§2.2"},{"comment":"The table header '∆NER' is used without defining whether the delta is absolute or relative; the text says '+79%/+73% relative' but the table shows absolute differences. Please clarify consistently.","section":"Table 2"},{"comment":"Equation (2) uses z_t both as a function and as the logit vector; this is a notation clash that could be clarified.","section":"§3.4"},{"comment":"The description of MCIF-dev says it is derived from MCIF, but the number of talks (21) and segments (919) is only given later in Table 6; consider stating these upfront.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The circularity of the primary terminology metrics and the unexplained GoldTag decrease on En→De are the core issues. The framework itself is interesting and the control analyses are thoughtful, but the paper's headline claim of consistent terminology-recall gains is not currently supported by an independent, reference-derived metric. I believe this is fixable within the manuscript's scope by making reference-supported matching the primary terminology result and explicitly reconciling the GoldTag discrepancy."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline claim — that extra document context in technical-talk SimulST should be converted into a sparse, per-segment terminology memory rather than injected uniformly — is a real and modest contribution, and the paper mostly earns it. The EGTA mechanism itself is simple: build a paper terminology list offline, match it against the streaming hypothesis, rank matches, then expose only the selected few terms through ASR hotwords/speech-side conditioning and decoder logit bias. No full-model fine-tuning. That is genuinely different from document-context MT and from global contextual biasing, and the paper tests it carefully.\n\nWhat impresses me is the design discipline. Shared backbone across variants, matched latency operating points, uniform-context and Global-G stress baselines, shuffled-memory control, segment- and talk-level bootstrap, and external ACL60/60-dev runs with fixed configuration. The shuffled-memory control is the most important number in the paper: on En→De the anchor NER gain drops from +0.266 with the correct memory to +0.012 with a shuffled memory. That tells me the effect is tied to the right paper's terms, not generic prompting. The reference-supported matching in §A.8 also shows large gains (micro-F1 from 0.625 to 0.839 on En-Zh, 0.575 to 0.753 on En-De), under a stricter criterion.\n\nNow the soft spots, in proportion. The primary NERclean/acronym metrics score outputs against gold terms produced by the same Qwen3-30B extractor that builds the memory. That makes the headline +79%/+73% relative NER gains partly a self-consistency measure, and the paper never states the gold is independently annotated. This is the main weakness. It does not sink the paper — the reference-supported diagnostic and shuffled-memory control survive it — but the abstract and §5.1 should lead with the reference-grounded numbers, not the self-referential ones. Second, \"consistent terminology-recall gains\" on ACL60/60-dev overstates Table 9: official GoldTag recall on En→De actually drops from 0.656 to 0.642. That is a small, fixable reporting problem, but as written it is a misdescription. Third, B=2 is tuned on the main benchmark; that should be acknowledged as a selected operating point. No code or checkpoints are released, which matters for a method whose appeal is being inference-time and cheap.\n\nBottom line: this is a serious paper with a real idea and unusually good controls. A proper referee would come back with a request to rework the metric story, not a rejection. I would send it to peer review.","headline":"A genuinely useful selective-terminology method for simultaneous speech translation, with unusually careful controls, but the headline terminology metric is partly self-referential and one external GoldTag cell drops; worth a serious referee, not a clean accept.","tokens_in":15065,"tokens_out":2464,"would_cite":true,"duration_ms":29679,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"In simultaneous translation of technical talks, the paper argues that extra document context is best used as sparse, evidence-triggered terminology adaptation rather than uniform context injection.","keywords":["simultaneous speech translation","terminology adaptation","document context","evidence grounding","streaming translation","technical talks","decoder biasing","term recall"],"falsifier":"Compute NER and acronym recall for both the baseline and EGTA-RG against an independently annotated gold set of paper terms per talk (for example, terms marked by human annotators or taken from the paper's own author glossary), using the same normalization. If the independent-gold gain is close to the paper's reference-supported numbers rather than the headline +0.212/+0.266, the evidence-conditioning claim is mostly self-consistency; if the gain persists, the claim is confirmed.","tokens_in":13945,"feed_emoji":"🎧","tokens_out":9286,"duration_ms":89438,"temperature":0.7,"pith_summary":"This paper is trying to establish a specific design principle for simultaneous speech translation of technical talks: extra document context (title, abstract, metadata) helps almost entirely because it lets the system recover paper-specific terminology, not because it provides general semantic enrichment. The paper therefore proposes replacing uniform full-context injection with a sparse, evidence-conditioned procedure — build a paper terminology memory, select only the terms the current streaming segment's partial speech or transcript actually matches, and push those selected terms through speech-side conditioning and decoder-side token biasing. On its technical-talk evaluation suite, this combined recipe improves term recall and acronym recall, maintains or improves a learned quality score and BLEU, and slightly reduces latency; the pattern transfers to a second benchmark without fine-tuning. If the claim is right, the practical lesson for streaming translation is that context should be treated as a selective, locally triggered term list rather than a uniform prompt.","feed_headline":"Sparse term lists beat full-document context in live translation","feed_subtitle":"Streaming technical-talk translation gains most from speech-triggered paper terms, not uniform context.","key_machinery":"The central object is the evidence-conditioned terminology memory: offline, document context is converted into a normalized list of paper-specific terms (acronyms, model and dataset names, multi-word technical phrases), treated as a noisy inventory rather than gold. Online, a deterministic selector tests each streaming segment's partial hypothesis, transcript history, and recent context against the memory, using surface-form, acronym, and tokenizer-variant matches, and ranks the hits to select a top-K candidate set. That set is then routed through two lightweight decision-space interfaces: R, which conditions the ASR or speech-understanding side (as hotwords or a compact prompt block), and G","core_discovery":"The paper's central claim is that the measurable benefit of document context in technical-talk simultaneous translation is concentrated in terminology recovery, and the right way to deliver it is a sparse, evidence-triggered term list rather than the full document. The proposed framework, EGTA, builds a terminology memory from paper-level metadata using a prompt-based LLM extractor, then for each streaming segment deterministically matches the current partial hypothesis or transcript to that memory, ranks matches by exactness, specificity, and recency, and keeps a compact top-K set. The selected set is applied through two interfaces: speech-side conditioning (hotwords or compact context) and","pith_inferences":["Editorial inference: because the terminology memory is built from the same extractor that later supplies the gold categories for the primary recall metric, the true external effect is likely closer to the paper's own reference-supported matching numbers than to the headline relative NER gains; an independent human-annotated gold set per segment would settle the gap.","Editorial inference: EGTA implies an upper bound set by the extractor — any paper term the memory misses can never be activated, so improving memory recall should directly raise achievable term recall; a testable extension is to run EGTA with deliberately degraded or enriched memories to map that dependency.","Editorial inference: the same sparse, evidence-conditioned activation design should transfer to other streaming settings with document-defined rare vocabulary, such as live subtitling of lectures, medical dictation, or meeting captioning, where the same rare-term failure mode appears."],"forward_implications":["Uniform full-document or full-memory context injection recovers only part of the terminology gain and can worsen the joint quality-latency trade-off.","Applying the selected terms on both the speech side and the decoder side gives larger and more consistent gains than either interface alone.","The gains are paper-specific: replacing the correct paper memory with a shuffled memory removes most of the terminology improvement, so the mechanism is evidence alignment rather than generic biasing.","Terminology recall improves across latency operating points without increasing latency, so the adaptation does not buy quality by waiting longer.","The same configuration transfers to a second technical-talk benchmark with no fine-tuning, indicating the finding is not an artifact of the primary test set."],"fun_headline_variants":["Terminology, not context, drives live translation gains","Sparse term memory boosts streaming translation","Evidence-grounded terms beat full doc context","Term-focused context wins for live translation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the headline terminology metrics measure externally valid term recall; in the paper the NERclean gold is derived from the same pipeline that produces the adaptation terms (the reference-supported matching is introduced separately as conservative), so if that gold is not independent, the large relative NER gains partly reflect the model echoing the extractor's vocabulary.","fun_headline_variants_meta":{"raw":{"variants":["Terminology, not context, drives live translation gains","Sparse term memory boosts streaming translation","Evidence-grounded terms beat full doc context","Term-focused context wins for live translation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000135,"raw_usage":{"total_tokens":1003,"prompt_tokens":790,"completion_tokens":213,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":158}},"tokens_in":534,"tokens_out":213,"duration_ms":3421,"temperature":1.0,"reasoning_tokens":158,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T17:06:09.307334+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute NER and acronym recall for both the baseline and EGTA-RG against an independently annotated gold set of paper terms per talk (for example, terms marked by human annotators or taken from the paper's own author glossary), using the same normalization. If the independent-gold gain is close to the paper's reference-supported numbers rather than the headline +0.212/+0.266, the evidence-conditioning claim is mostly self-consistency; if the gain persists, the claim is confirmed.","supporting_citations":[],"review_version":1}