MWE-ECL: 可恢复的长距离上下文未必总能覆盖局部词汇先验
摘要
本文介绍了MWE-ECL,一个诊断工具,用于测试AI模型是否能利用长距离上下文来覆盖多词表达中的局部语义先验,突出了可恢复性与行为影响之间的差距。
arXiv:2609.27590v1 Announce Type: new
Abstract: Long-context evaluations often test whether a model can recover distant evidence, but recoverability does not guarantee behavioral influence. We test the prediction that a distant discourse anchor can remain explicitly recoverable yet fail to change the locally preferred reading of a familiar multiword expression; such failures should concentrate when the model's no-anchor default conflicts with the anchor, while prior-correct decisions remain largely preserved. We introduce Multiword Expression Effective Context Length (MWE-ECL), a bilingual diagnostic whose matched anchor-retrieval, no-anchor prior, and interpretation prompts measure explicit recoverability, model-observed defaults, and anchor-conditioned decisions, respectively. Across eight English deployment panels on a shared 0-128K grid, retrieval-control accuracy on prior-conflict items is 0.989-1.000, prior-conflict override spans 0.806-1.000 (0.809-1.000 after conditioning on correct retrieval), and preservation of prior-correct decisions remains 0.977-1.000. A same-call control querying retrieval and interpretation in one prompt reproduces the gap for DeepSeek V4 Pro (1.000 retrieval versus 0.900-0.920 interpretation), showing that separate invocations are not its sole explanation; smaller or absent gaps in the other two models bound its generality. For DeepSeek V4 Flash, separate prompt-fit tests retain perfect retrieval with lower interpretation at 512K and 1M, while foil-consistent cues shift the no-anchor prior far more than retrieval; cross-model cue effects are heterogeneous. A separately reported 10-family Chinese subset shows similar descriptive gaps, but imperfect retrieval for some models prevents an integration-only attribution. MWE-ECL therefore evaluates whether explicitly recoverable distant context changes a competing local semantic decision.
查看缓存全文
缓存时间: 2026/09/24 09:25
# MWE-ECL: Recoverable Long-Range Context Does Not Always Override Local Lexical Priors
Source: [https://arxiv.org/html/2609.27590](https://arxiv.org/html/2609.27590)
###### Abstract
Long\-context evaluations often test whether a model can recover distant evidence, but recoverability does not guarantee behavioral influence\. We test the prediction that a distant discourse anchor can remain explicitly recoverable yet fail to change the locally preferred reading of a familiar multiword expression; such failures should concentrate when the model’s no\-anchor default conflicts with the anchor, while prior\-correct decisions remain largely preserved\. We introduce Multiword Expression Effective Context Length \(MWE\-ECL\), a bilingual diagnostic whose matched anchor\-retrieval, no\-anchor prior, and interpretation prompts measure explicit recoverability, model\-observed defaults, and anchor\-conditioned decisions, respectively\. Across eight English deployment panels on a shared 0–128K grid, retrieval\-control accuracy on prior\-conflict items is 0\.989–1\.000, prior\-conflict override spans 0\.806–1\.000 \(0\.809–1\.000 after conditioning on correct retrieval\), and preservation of prior\-correct decisions remains 0\.977–1\.000\. A same\-call control querying retrieval and interpretation in one prompt reproduces the gap for DeepSeek V4 Pro \(1\.000 retrieval versus 0\.900–0\.920 interpretation\), showing that separate invocations are not its sole explanation; smaller or absent gaps in the other two models bound its generality\. For DeepSeek V4 Flash, separate prompt\-fit tests retain perfect retrieval with lower interpretation at 512K and 1M, while foil\-consistent cues shift the no\-anchor prior far more than retrieval; cross\-model cue effects are heterogeneous\. A separately reported 10\-family Chinese subset shows similar descriptive gaps, but imperfect retrieval for some models prevents an integration\-only attribution\. MWE\-ECL therefore evaluates whether explicitly recoverable distant context changes a competing local semantic decision\.
1University of Exeter
2University of Cambridge
## 1Introduction
A model may correctly report a constraint stated more than 100,000 tokens earlier yet make the same later decision it would have made without that constraint\. A million\-token context window specifies nominal capacity, not a guarantee of effective use\. Prior work operationalizes effective context length \(ECL\) as the largest tested context length at which task performance remains satisfactory\([Hsieh et al\. 2024](https://arxiv.org/html/2609.27590#bib.bib5);[An et al\. 2025](https://arxiv.org/html/2609.27590#bib.bib6)\)\. Here, we focus on whether recoverable distant evidence changes a decision when it conflicts with a strong local expectation\.
This distinction matters for long\-running assistants and agents, where a short current turn may depend on a preference, policy exception, or state update established far earlier\.
Retrieval suites reveal position\-dependent access failures\([Liu et al\. 2024](https://arxiv.org/html/2609.27590#bib.bib4);[Hsieh et al\. 2024](https://arxiv.org/html/2609.27590#bib.bib5)\), while broader evaluations show that exact recoverability need not ensure use\([Goldman et al\. 2024](https://arxiv.org/html/2609.27590#bib.bib9);[Yang 2024](https://arxiv.org/html/2609.27590#bib.bib14);[Du et al\. 2025](https://arxiv.org/html/2609.27590#bib.bib8);[Chen et al\. 2026](https://arxiv.org/html/2609.27590#bib.bib15)\)\. Multiword Expression Effective Context Length \(MWE\-ECL\) operationalizes this distinction as lexical\-semantic prior override\. Familiar multiword expressions \(MWEs\) invite conventional local readings, but a distant discourse anchor can license the competing sense\([Haagsma et al\. 2020](https://arxiv.org/html/2609.27590#bib.bib22);[Mi et al\. 2025](https://arxiv.org/html/2609.27590#bib.bib23)\)\. The model must therefore assess whether to use a distant but accessible state to change a local decision\. In our design, increasing neutral filler jointly increases the total context size and the distance between the relevant anchor and the decision point\.
MWEs provide a controlled version of this conflict\. For example,*spill the beans*normally favors a figurative reading, whereas a distant warehouse scene can license a literal one\. The local trigger is identical across senses, so lexical familiarity cannot solve the task; it is the anchor that alters the decision\. This isolates prior override without claiming to reproduce every component of an agent workflow\.
Each literal/figurative case yields matched interpretation, anchor\-retrieval, and no\-anchor prior\-control prompts \(Figure[1](https://arxiv.org/html/2609.27590#S1.F1)\)\. These heads separate explicit anchor access, the model\-observed local default, and the anchored decision\. ECL is reported only as right\-censored tested\-range coverage; prior\-conflict override is the targeted context\-use metric\.
We ask three research questions\.RQ1 \(Access\):Can models identify the distant anchor when asked directly across the tested distances?RQ2 \(Override\):When the anchor\-free prompt elicits the wrong interpretation, does the matched anchor\-present prompt override that choice?RQ3 \(Robustness\):Does the access–use gap persist at longer distances, under semantic interference, in a second language, and within a same\-call control? This paper makes three contributions
Figure 1:MWE\-ECL prompt timeline and matched controls\. The heads separate \(1\) anchor\-conditioned interpretation, \(2\) explicit recoverability, and \(3\) the model\-observed default; high retrieval with lower interpretation indicates insufficient contextual influence at the behavioral level\.- •We introduce MWE\-ECL, a bilingual long\-context diagnostic whose matched interpretation, retrieval\-control, and no\-anchor prior\-control heads separate explicit anchor recoverability, model\-observed local defaults, and anchor\-conditioned decisions\.
- •We define prior\-conflict override and add retrieval\-conditioned override and preservation, distinguishing correction of a wrong local default from access and corruption of an already correct default\.
- •We reveal a recurring behavioral access–use dissociation in English across eight deployment panels and context scales, with descriptive supporting evidence under semantic interference and in Chinese\.
## 2Related Work
#### Long\-context access and use\.
LongBench, RULER, NeedleBench, LongBench v2, and NoLiMa progressively broaden long\-context access, aggregation, reasoning, and low\-overlap retrieval\([Bai et al\. 2024](https://arxiv.org/html/2609.27590#bib.bib18);[Hsieh et al\. 2024](https://arxiv.org/html/2609.27590#bib.bib5);[Li et al\. 2025](https://arxiv.org/html/2609.27590#bib.bib19);[Bai et al\. 2025](https://arxiv.org/html/2609.27590#bib.bib20);[Modarressi et al\. 2025](https://arxiv.org/html/2609.27590#bib.bib7)\)\. Efficient inference methods such as DuoAttention address the cost of serving long contexts\([Xiao et al\. 2025](https://arxiv.org/html/2609.27590#bib.bib3)\); our concern is complementary: whether served context changes behavior\. These suites establish that nominal context length overstates effective access: position effects remain substantial\([Liu et al\. 2024](https://arxiv.org/html/2609.27590#bib.bib4)\), and retrieval becomes harder as lexical overlap and evidence locality decrease\. Yet a successful lookup still leaves open whether the recovered state influences a downstream choice\. MWE\-ECL therefore places an explicit retrieval control beside a decision whose correct answer can oppose a strong local lexical default\.
#### Beyond retrieval: integration and knowledge conflict\.
Beyond\-retrieval benchmarks distinguish localized access from holistic or dispersed\-evidence integration\([Goldman et al\. 2024](https://arxiv.org/html/2609.27590#bib.bib9);[Yang 2024](https://arxiv.org/html/2609.27590#bib.bib14);[Du et al\. 2025](https://arxiv.org/html/2609.27590#bib.bib8);[Chen et al\. 2026](https://arxiv.org/html/2609.27590#bib.bib15)\)\. Knowledge\-conflict work likewise shows that accessible context can lose to memorized associations\([Xie et al\. 2024](https://arxiv.org/html/2609.27590#bib.bib17);[Du et al\. 2024](https://arxiv.org/html/2609.27590#bib.bib10);[Sun et al\. 2026](https://arxiv.org/html/2609.27590#bib.bib11);[Fu et al\. 2026](https://arxiv.org/html/2609.27590#bib.bib16)\)\. These lines of work motivate a stricter question than fact recovery: whether context changes a response when an alternative is already supported by parametric or local evidence\. MWE\-ECL contributes a controlled lexical\-semantic version of this problem, rather than the general retrieval–use distinction itself\. Its no\-anchor head measures the competing default on the same trigger, and its retrieval head verifies that the distant scenario remains explicitly recoverable without requiring the model to solve the semantic conflict\.
#### Retrieval\-Augmented Generation and Agent Memory\.
Retrieval\-augmented generation \(RAG\) and memory evaluations include retriever quality, faithfulness, temporal reasoning, and updates\([Hagström et al\. 2025](https://arxiv.org/html/2609.27590#bib.bib27);[Wu et al\. 2025](https://arxiv.org/html/2609.27590#bib.bib28)\)\. MWE\-ECL removes the external retriever: the anchor is already in context, eliminating document selection as a source of failure\. It also differs from answer\-bearing RAG evaluation because the anchor does not simply contain the final answer; it licenses one reading of a later ambiguous expression\. The test is therefore whether available memory changes an activated interpretation, a pattern that also arises when agents must apply an old preference or exception to a new turn\.
#### Multiword expressions\.
MWEs are non\-compositional and context\-sensitive\([Sag et al\. 2002](https://arxiv.org/html/2609.27590#bib.bib1);[Baldwin and Kim 2010](https://arxiv.org/html/2609.27590#bib.bib2);[Constant et al\. 2017](https://arxiv.org/html/2609.27590#bib.bib21)\)\. MAGPIE, DICE, PaCE, and MIDI support literal/idiomatic and multilingual analysis\([Haagsma et al\. 2020](https://arxiv.org/html/2609.27590#bib.bib22);[Mi et al\. 2025](https://arxiv.org/html/2609.27590#bib.bib23);[Li et al\. 2026](https://arxiv.org/html/2609.27590#bib.bib12);[Almheiri et al\. 2026](https://arxiv.org/html/2609.27590#bib.bib13)\); prior probes are usually sentence\- or short\-context based and primarily assess sense classification or contextual bias\. Recent work investigates idiomaticity in word representations\([He et al\. 2025](https://arxiv.org/html/2609.27590#bib.bib25)\); complementary multimodal work uses semantic anchorage to quantify a semiotic gap\([He 2026](https://arxiv.org/html/2609.27590#bib.bib26)\)\. MWE\-ECL instead separates a visible phrase from its licensing evidence, pairs literal and figurative cases within each family, and inserts controlled long\-distance filler\. Matched retrieval and model\-specific prior controls then distinguish failure to access the licensing scenario from failure to use it\.
## 3Benchmark Design
MWE\-ECL is a diagnostic stress test, not a benchmark of average\-case language understanding\. It intentionally selects familiar expressions for which a conventional local reading competes with an alternative interpretation licensed by distant discourse\. Its matched interpretation, anchor\-retrieval, and no\-anchor prior\-control heads separate the anchor\-conditioned decision, explicit recoverability, and the model\-observed default\. Jointly low retrieval and interpretation is compatible with access failure; high retrieval with lower interpretation indicates insufficient contextual influence; and low preservation indicates anchor\-induced corruption\. These are behavioral, not hidden\-state, conclusions\.
### 3\.1Task Structure
Each anchor\-present prompt contains a start\-positionanchor, deterministic neutralfiller, an end\-position ambiguoustrigger, and a binaryquestion\. All eight primary panels use nominal 0/512/2K/8K/32K/64K/128K filler\. K denotes 1,024 tokens\. Start placement and an explicit whole\-document instruction favor access; a small anchor\-position pilot tests this placement choice\.
The construction holds the local decision surface fixed while varying the distant licensing context\. Within each family, literal and figurative variants share the same trigger sentence and answer options; the distant anchor determines which interpretation is licensed, while the prior control removes that anchor\. Target\-independent filler increases both context size and anchor–decision distance without adding sense\-relevant evidence, so the primary grid does not disentangle their effects\. This design trades document naturalism for a controlled contrast between access and use\.
Table 1:Matched zero\-filler example\. Opposing anchors reverse the reference answer for a shared trigger and option set; anchors are abbreviated, and longer records insert neutral filler\.Table[1](https://arxiv.org/html/2609.27590#S3.T1)illustrates a paired family: two distant anchors license opposite readings of the same*spill the beans*trigger and option set\. The forced\-choice format eliminates the need for a judge model; anchors are manually contrastive, and filler contains neither the target MWE nor either licensing scenario\.
Because provider tokenization differs, nominal 128K filler yields approximately 136,671 and 153,552 prompt tokens on representative DeepSeek and Qwen endpoints, respectively\. We therefore use nominal filler distance for cross\-model display and retain exact prompt\-token counts in the reproducibility audit\.
### 3\.2Controls and Hard\-Cue Extension
The retrieval control asks for anchor\-scenario identification rather than MWE interpretation\. The no\-anchor prior control replaces the anchor with a non\-disambiguating note while retaining the trigger and answer options, yielding an operational model default rather than a human lexical prior\. Because the main heads use separate calls, their difference is a behavioral between\-call contrast; an order\-balanced same\-call control tests whether the retrieval–interpretation gap persists within one invocation\.
The hard\-cue extension replaces neutral filler with target\-consistent cues supporting the anchor\-licensed interpretation, foil\-consistent cues supporting its competitor, or mixed cues containing both, while holding the anchor, trigger, and answer options fixed\.
### 3\.3Metrics
For modelmm, tasktt, and distancedd, letRm,t,dR\_\{m,t,d\}denote the matched case–repetition identities for which taskttyields a non\-error prediction\. Accuracy is computed over these task\-specific identities:
Acc\(m,t,d\)=1\|Rm,t,d\|∑r∈Rm,t,d𝟏\{y^r=yr\}\.\\mathrm\{Acc\}\(m,t,d\)=\\frac\{1\}\{\|R\_\{m,t,d\}\|\}\\sum\_\{r\\in R\_\{m,t,d\}\}\\mathbf\{1\}\\\{\\hat\{y\}\_\{r\}=y\_\{r\}\\\}\.\(1\)Here,yry\_\{r\}is the gold label for identityrr, andy^r\\hat\{y\}\_\{r\}is the model prediction\. At thresholdτ=0\.8\\tau=0\.8, joint ECL is the furthest tested distance where both interpretation and retrieval pass\. Because distance cells are independently counterbalanced and non\-monotone, ECL is right\-censored coverage, not an estimated failure boundary\. Confidence intervals \(CIs\) use 10,000 whole\-family bootstrap resamples; case\-clustered estimates provide a sensitivity check\.
#### Override rate\.
For the comparisons below, letRm,d=Rm,prior,d∩Rm,int,dR\_\{m,d\}=R\_\{m,\\mathrm\{prior\},d\}\\cap R\_\{m,\\mathrm\{int\},d\}denote the matched identities with non\-error prior and interpretation predictions\. We define the*override rate*on the model\-specific prior\-conflict set:
𝒞m,d=\{r∈Rm,d:y^rprior≠yr\}\.\\mathcal\{C\}\_\{m,d\}=\\\{\\,r\\in R\_\{m,d\}:\\hat\{y\}^\{\\mathrm\{prior\}\}\_\{r\}\\neq y\_\{r\}\\,\\\}\.\(2\)OR\(m,d\)=1\|𝒞m,d\|∑r∈𝒞m,d𝟏\{y^rint=yr\}\.\\mathrm\{OR\}\(m,d\)=\\frac\{1\}\{\|\\mathcal\{C\}\_\{m,d\}\|\}\\sum\_\{r\\in\\mathcal\{C\}\_\{m,d\}\}\\mathbf\{1\}\\\{\\hat\{y\}^\{\\mathrm\{int\}\}\_\{r\}=y\_\{r\}\\\}\.\(3\)Here,y^rprior\\hat\{y\}^\{\\mathrm\{prior\}\}\_\{r\}andy^rint\\hat\{y\}^\{\\mathrm\{int\}\}\_\{r\}denote the no\-anchor prior\-control and interpretation predictions, respectively\. OR is the fraction of prior\-conflict identities corrected once the anchor is present\.
Aggregate OR pools its numerator and denominator over all tested distancesDmD\_\{m\}in modelmm’s panel:
OR\(m\)=∑d∈Dm∑r∈𝒞m,d𝟏\{y^intr=yr\}∑d∈Dm\|𝒞m,d\|\.\\mathrm\{OR\}\(m\)=\\frac\{\\sum\_\{d\\in D\_\{m\}\}\\sum\_\{r\\in\\mathcal\{C\}\_\{m,d\}\}\\mathbf\{1\}\\\{\\hat\{y\}^\{\\mathrm\{int\}\}\_\{r\}=y\_\{r\}\\\}\}\{\\sum\_\{d\\in D\_\{m\}\}\|\\mathcal\{C\}\_\{m,d\}\|\}\.\(4\)
To condition on retrieval success, let𝒞m,dret=𝒞m,d∩Rm,ret,d\\mathcal\{C\}^\{\\mathrm\{ret\}\}\_\{m,d\}=\\mathcal\{C\}\_\{m,d\}\\cap R\_\{m,\\mathrm\{ret\},d\}denote prior\-conflict identities with a non\-error retrieval prediction\. We report matched retrieval\-conditioned override as
OR\|Ret\(m,d\)=∑r∈𝒞m,dret𝟏\{y^retr=yr∧y^intr=yr\}∑r∈𝒞m,dret𝟏\{y^retr=yr\}\.\\mathrm\{OR\{\\mid\}Ret\}\(m,d\)=\\frac\{\\sum\_\{r\\in\\mathcal\{C\}^\{\\mathrm\{ret\}\}\_\{m,d\}\}\\mathbf\{1\}\\\{\\hat\{y\}^\{\\mathrm\{ret\}\}\_\{r\}=y\_\{r\}\\land\\hat\{y\}^\{\\mathrm\{int\}\}\_\{r\}=y\_\{r\}\\\}\}\{\\sum\_\{r\\in\\mathcal\{C\}^\{\\mathrm\{ret\}\}\_\{m,d\}\}\\mathbf\{1\}\\\{\\hat\{y\}^\{\\mathrm\{ret\}\}\_\{r\}=y\_\{r\}\\\}\}\.\(5\)Here,y^rret\\hat\{y\}^\{\\mathrm\{ret\}\}\_\{r\}denotes the retrieval\-control prediction\. OR∣\\midRet conditions on success in a separate matched call; it does not prove that the interpretation call retrieved the anchor\.
Correction is only one possible anchor effect\. We therefore define
Pres\(m,d\)\\displaystyle\\mathrm\{Pres\}\(m,d\)=1\|Rm,d∖𝒞m,d\|∑r∈Rm,d∖𝒞m,d𝟏\{y^intr=yr\},\\displaystyle=\\frac\{1\}\{\|R\_\{m,d\}\\setminus\\mathcal\{C\}\_\{m,d\}\|\}\\sum\_\{r\\in R\_\{m,d\}\\setminus\\mathcal\{C\}\_\{m,d\}\}\\mathbf\{1\}\\\{\\hat\{y\}^\{\\mathrm\{int\}\}\_\{r\}=y\_\{r\}\\\},\(6\)Corrup\(m,d\)\\displaystyle\\mathrm\{Corrup\}\(m,d\)=1−Pres\(m,d\)\.\\displaystyle=1\-\\mathrm\{Pres\}\(m,d\)\.The matched\-valid setRm,dR\_\{m,d\}partitions into prior\-conflict and prior\-correct identities; OR and preservation summarize interpretation correctness on these two subsets, respectively\. Corruption is the complement of preservation, where the anchor changes an already\-correct default into an incorrect interpretation\.
Aggregate OR∣\\midRet and preservation entries likewise pool their corresponding numerators and denominators overd∈Dmd\\in D\_\{m\}; they are not unweighted averages of per\-distance rates\. Rates with zero denominators are reported as NA, and every nonzero OR∣\\midRet denominator is reported\.
These conditional rates are model\-specific by design\. Whenever𝒞m,d\\mathcal\{C\}\_\{m,d\}is nonempty, a model that always follows its no\-anchor default has an override rate of zero regardless of overall accuracy\. For cross\-model comparability, we additionally report conflict\-set overlap and results on a fixed consensus\-hard subset in the supplementary conflict\-set tables\. The supplementary hard\-cue tables also report integration under retrieval \(IUR\), defined as interpretation accuracy conditioned on correct anchor retrieval over the same identities\.
## 4Dataset
The bilingual core comprises an English \(EN\) subset of 50 cases over 25 MWE families and a Simplified Chinese \(ZH\) subset of 20 cases over 10 families, with one literal and one figurative case per family \(Table[2](https://arxiv.org/html/2609.27590#S4.T2)\)\. The English cases span four construction types; for 38 English cases, the target expressions were cross\-checked against DICE\([Mi et al\. 2025](https://arxiv.org/html/2609.27590#bib.bib23)\)and MAGPIE\([Haagsma et al\. 2020](https://arxiv.org/html/2609.27590#bib.bib22)\)\. All benchmark anchors, triggers, questions, and options are newly authored\.
Table 2:Dataset composition\. Each family supplies one literal and one figurative case\. Family mix is English \(EN\): 14 idiom, 8 compound, 2 prepositional, 1 verb\-particle; Chinese \(ZH\): 4 idiom, 4 compound, 1 verb\-object, 1 domain\-derived\. The three\-distance expansion is automatically checked, not human\-validated or pooled\.The idiom\-heavy English catalog intentionally stresses cases with strong figurative defaults, so results may therefore not generalize to constructions with weaker or domain\-specific priors\. The separate 60\-case English expansion broadens coverage over 30 families at 0/8K/64K, but is not difficulty\-matched, human\-validated, pooled, or used for ranking\.
The Chinese subset selects 10 families from XMPIE\([Torunoğlu\-Selamet et al\. 2026](https://arxiv.org/html/2609.27590#bib.bib24)\)and retains only source expressions and sense evidence; no source sentence is copied\. It uses the same three tasks at seven distances through 128K, but its 20 cases are reported separately because difficulty is not matched across languages\. Distances are approximate under provider tokenization, and prompt\-token counts are recorded when available\.
Automated checks cover schema, class balance, provenance, prompt construction, filler generation, and static leakage; filler contains neither verbatim answer options nor target MWEs\. Two English\-proficient annotators and two native\-Chinese annotators, all blind to model outputs, independently selected the anchor\-licensed labels\. They matched all intended labels \(100/100 English judgments and 40/40 Chinese judgments\) and flagged no items\.
## 5Experiments
### 5\.1Models
Eight application programming interface \(API\) deployment panels form the headline comparison: DeepSeek V4 Flash/Pro\([DeepSeek\-AI 2026](https://arxiv.org/html/2609.27590#bib.bib29)\), MiniMax\-M3\([MiniMax 2026](https://arxiv.org/html/2609.27590#bib.bib30)\), Qwen3\.6\-Plus and Qwen3\.6\-35B\-A3B\([Alibaba Cloud 2026](https://arxiv.org/html/2609.27590#bib.bib31);[Qwen Team 2026](https://arxiv.org/html/2609.27590#bib.bib32)\), GLM\-4\.7 and GLM\-5\.1\([Zhipu AI 2026a](https://arxiv.org/html/2609.27590#bib.bib33);[Zhipu AI 2026b](https://arxiv.org/html/2609.27590#bib.bib34)\), and Gemma\-4\-26B\-A4B\-it\([Gemma Team 2026](https://arxiv.org/html/2609.27590#bib.bib35)\)\. All eight models complete the same seven\-distance 0–128K English grid \(1,050 records/model\) and the corresponding Chinese grid \(420 records/model\), with the language subsets reported separately\. Comparisons across mismatched grids or endpoint configurations are descriptive deployment snapshots, not architecture rankings\.
Separate stress regimes evaluate Qwen3\.6\-35B\-A3B and Gemma\-4\-26B\-A4B\-it near 256K, and DeepSeek V4 Flash at 512K and 1M\. Hard\-cue tests include a 3,300\-row DeepSeek V4 Flash study spanning neutral, target\-consistent, foil\-consistent, and mixed conditions, and all\-family neutral/target/foil cue triads for six additional models at 8K and 64K; matched pilot slices cover DeepSeek V4 Flash and GLM\-5\.1\.
### 5\.2Implementation
Each chat\-completion prompt instructs the model to use the entire document and answer with one letter\. Runs use temperature00and a four\-token answer budget\. The first explicitA/Bis extracted; completed but unparsed outputs are incorrect, whereas provider failures are excluded and reported separately\.
Five panels disable extended thinking; Qwen3\.6\-Plus and both GLMs use provider defaults\. Thus, failures occur under explicit context\-use\-favoring instructions\.
### 5\.3Evaluation Protocol
All ECL calculations useτ=0\.8\\tau=0\.8, and confidence intervals use the stated whole\-family bootstrap procedure\. Gold A/B labels are balanced within every task–distance cell\. Counterbalancing may swap a case’s option order across distances, so cells are coverage snapshots, not longitudinal trajectories\. Matched\-grid, prompt\-fit, hard\-cue, and Chinese evaluations remain separate whenever grids, filler semantics, or languages differ\.
Repeated\-run checks cover Qwen3\.6\-35B\-A3B and Gemma\-4\-26B\-A4B\-it at 64K and in three 128K passes; same\-call controls cover DeepSeek V4 Pro, GLM\-5\.1, and MiniMax\-M3 with retrieval and interpretation heads in the same call at 64K/128K\. Other headline panels are single\-pass; bootstrap intervals quantify item/family variation, not API nondeterminism\.
Primary within\-model dissociation tests compare retrieval\-control accuracy and override on identical conflict rows\. We use exact family\-clustered sign\-flip tests with Holm correction over eight pre\-specified within\-model tests; effect sizes and whole\-family bootstrap CIs remain primary\. Cross\-model tests are secondary because model\-specific conflict sets, grids, and endpoint configurations can differ\.
The shared 0–128K EN grid is the primary evaluation\. Same\-call, hard\-cue, prompt\-fit, expansion, and ZH analyses probe invocation separation, semantic interference, longer\-context stress, lexical coverage, and cross\-language robustness, respectively; these regimes are not pooled into one ranking\.
## 6Results
Figure 2:Core English 0–128K results\.Left:retrieval versus OR∣\\midRet on prior\-conflict rows\.Right:interpretation by distance\. Three models share the ceiling; bars/bands are family\-bootstrap 95% CIs and cells are counterbalanced snapshots\.### 6\.1Near\-Ceiling Recoverability Does Not Ensure Prior\-Conflict Correction
RQ1–RQ2\.On shared\-grid prior\-conflict rows, retrieval is 0\.989–1\.000, whereas OR and OR∣\\midRet span 0\.806–1\.000 and 0\.809–1\.000 \(Figure[2](https://arxiv.org/html/2609.27590#S6.F2)\)\. DeepSeek V4 Pro retrieves at 0\.989 but attains 0\.806 OR over 175 conflicts \(gap 0\.183, 95% family\-bootstrap CI\[0\.088,0\.295\]\[0\.088,0\.295\]; exact family sign\-flipp<0\.001p<0\.001, Holmp=0\.002p=0\.002\)\. Its OR∣\\midRet is 0\.809 \(n=173n=173\{\}\)\. MiniMax\-M3’s gap is 0\.089 \(\[0\.038,0\.148\]\[0\.038,0\.148\]; Holmp=0\.055p=0\.055; conflictn=180n=180\)\. Thus recoverable evidence does not always change the locally preferred interpretation\.
Table 3:English and Chinese subset results \(1,050/420 records per model\), reported separately\.CCandnndenote the OR and OR∣\\midRet denominators; Pres\. is prior\-correct preservation\. The subsets are not pooled or difficulty\-matched\.Table[3](https://arxiv.org/html/2609.27590#S6.T3)shows near\-ceiling retrieval but lower interpretation and override in several panels\. All eight panels pass at the 128K endpoint underτ=0\.8\\tau=0\.8, so joint ECL is right\-censored rather than a model ranking\. Atτ=0\.9\\tau=0\.9, joint ECL drops to 512 tokens for DeepSeek V4 Pro and 64K for MiniMax\-M3; the other six panels remain right\-censored at 128K\.
Aggregate prior\-control accuracy against the anchor\-licensed labels is 0\.486–0\.514, while semantic agreement between 0K and 128K no\-anchor choices, after mapping A/B to literal/figurative senses, is 0\.820–0\.960\. Thus, the near\-0\.5 aggregate does not imply random or wholly unstable defaults\.
At 128K, interpretation is 0\.940/0\.880/0\.860 for DeepSeek V4 Flash/Pro and MiniMax\-M3; Qwen3\.6\-35B/Gemma\-4\-26B reach 0\.940/0\.900, and Qwen3\.6\-Plus, GLM\-4\.7, and GLM\-5\.1 remain at 1\.000\.
Balanced labels and same\-order checks reduce simple option\-position explanations: on 146 shared 0–64K rows, DeepSeek V4 Pro still reaches 0\.993 retrieval but 0\.890 interpretation and 0\.768 override \(69 conflicts\)\.
### 6\.2The Gap Is Not Solely a Separate\-Call Artifact
RQ3\.A targeted same\-call control removes invocation separation by asking both heads within one shared prompt\. Each model completes 200 rows over 50 cases, two distances, and two counterbalanced query orders \(Table[4](https://arxiv.org/html/2609.27590#S6.T4)\)\.
Table 4:Same\-call dual\-head control\. Ranges are minima and maxima across the four distance\-by\-query\-order cells; Gap is retrieval minus interpretation\. Because the prior head requires an anchor\-free prompt, this control does not estimate override rate\.Order effects are small and their family\-bootstrap CIs include zero\. DeepSeek V4 Pro nevertheless makes four or five interpretation errors in each 50\-case cell while retrieving all 50 anchors correctly\. Separate invocations therefore cannot be the sole source of its gap; the GLM\-5\.1 ceiling result also bounds the claim’s generality\.
### 6\.3Failures Concentrate Where Local Defaults Conflict
RQ2\.Because the suite is sense\-balanced, aggregate prior\-control accuracy near 0\.5 can hide a directional figurative default\. OR therefore conditions on each model’s own prior errors\. Preservation is 0\.977–1\.000, indicating that anchors mostly correct prior conflicts rather than disrupt already\-correct defaults: DeepSeek V4 Pro corrupts 4 of 175 prior\-correct rows, MiniMax\-M3 2 of 170, and Qwen3\.6\-35B 1 of 173; the other panels corrupt none\.
Shared\-row and fixed\-conflict analyses show the same pattern\. On the consensus\-hard 0–128K primary subset \(n=182n\{=\}182\{\}\), fixed\-subset interpretation accuracies are 0\.940/0\.813, versus 0\.918/1\.000 for MiniMax\-M3/GLM\-4\.7\. Literal\-minus\-figurative interpretation gaps are−0\.063\-0\.063for DeepSeek V4 Flash,−0\.137\-0\.137for Pro,−0\.091\-0\.091for Qwen3\.6\-35B, and−0\.074\-0\.074for Gemma\. The corresponding family\-bootstrap CIs are\[−0\.143,−0\.006\]\[\-0\.143,\-0\.006\],\[−0\.269,−0\.017\]\[\-0\.269,\-0\.017\],\[−0\.194,0\.011\]\[\-0\.194,0\.011\], and\[−0\.143,−0\.017\]\[\-0\.143,\-0\.017\], respectively; they exclude zero for Flash, Pro, and Gemma, but not Qwen\. Holm\-adjusted sign\-flip tests remain descriptive \(p≥0\.250p\\geq 0\.250\), so we claim directional concentration rather than a universal sense effect\. OR∣\\midRet sense effects are heterogeneous\.
All errors in the five highest\-miss families occur on literal variants \(Table[5](https://arxiv.org/html/2609.27590#S6.T5)\)\. For all 19*cold feet*interpretation\-error events, the matched retrieval prediction identifies the literal anchor, whereas both the no\-anchor prior and anchored interpretation select the figurative option\. This provides behavioral evidence that correct matched retrieval can coexist with a response aligned with the local default\.
Table 5:English families with the most interpretation errors across the eight 0–128K panels\. Counts aggregate model–distance records; all families tied at the cutoff are shown\.
### 6\.4Stress Tests Reproduce and Bound the Dissociation
RQ3\.Separate prompt\-fit snapshots retain 1\.000 retrieval while interpretation remains lower at the 256K, 512K, and 1M settings \(Table[6](https://arxiv.org/html/2609.27590#S6.T6)\)\. Because these runs use endpoint\-specific prompt\-fit configurations, they are not a monotone extension of the matched grid\.
Table 6:Prompt\-fit stress outside the matched grid\. Retrieval stays at ceiling while interpretation is lower\.With 4 and 16 foil\-consistent cues, DeepSeek V4 Flash interpretation falls to 0\.873 and 0\.867, respectively\. At 16 cues, the no\-anchor prior shifts far more than either retrieval or integration under retrieval \(Table[7](https://arxiv.org/html/2609.27590#S6.T7)\), supporting semantic interference rather than simple anchor loss\.
Table 7:Hard\-cue outcomes at 16 foil cues\. DeepSeek V4 Flash uses IUR across all 25 families; the six\-model panels use interpretation accuracy\. The final column reports across\-panel min–max ranges, not confidence intervals\. Deltas are foil minus neutral\.Cue effects are heterogeneous\. In the matched pilot, foil cues change interpretation by−0\.100\-0\.100for DeepSeek V4 Flash and\+0\.033\+0\.033for GLM\-5\.1\. In the all\-family runs, the corresponding changes are−0\.080\-0\.080for DeepSeek V4 Pro,−0\.060\-0\.060for Qwen3\.6\-Plus, and−0\.010\-0\.010to\+0\.020\+0\.020for the remaining four models\. Foil cues shift prior\-control behavior more consistently than interpretation or retrieval, whose effects remain deployment\-dependent\.
The 60\-case expansion also shows lower override than retrieval across four models \(override 0\.867–0\.897 vs\. retrieval 0\.961–0\.983\)\. Because it is neither human\-validated nor difficulty\-matched, it serves as a lexical\-coverage check rather than ranking evidence\.
The Chinese subset shows the same qualified pattern \(Table[3](https://arxiv.org/html/2609.27590#S6.T3)\): three models are at joint ceiling, while DeepSeek V4 Flash reaches 1\.000 retrieval/0\.986 OR and the other four non\-ceiling panels reach 0\.921–0\.986 retrieval/0\.771–0\.900 OR\. At 128K, Qwen3\.6\-35B and Gemma score 0\.950/0\.950 and 0\.900/0\.950 interpretation/retrieval\. Because retrieval itself is imperfect for some models and the subset has only 10 families, these results are descriptive, non\-pooled, and not a cross\-language difficulty comparison\. Preservation is 1\.000 in seven panels, whereas DeepSeek V4 Pro flips 8 of 70 prior\-correct rows \(Pres\. 0\.886\); its Chinese gap therefore combines weaker correction with anchor\-induced corruption\.
A 20\-item Qwen3\.6\-35B\-A3B position pilot holds the trigger at the end\. At 32K, start yields IUR/OR of 0\.95/0\.90 versus 0\.85/0\.70 at middle and late positions\. At 64K, the corresponding values are 0\.85/0\.70 versus 0\.84/0\.60; each non\-start IUR estimate conditions on 19 retrieval\-correct rows\. At 8K, start and late yield 0\.95/0\.90 versus 0\.85/0\.70 at middle\. The small\-NNpilot therefore does not rule out a start\-position advantage\.
The evidence instead supports competition from a local lexical default: literal errors concentrate in conventional idioms, hard cues shift the prior more than retrieval, and*cold feet*rows retrieve the literal anchor while still selecting the figurative interpretation\. This parallels contextual knowledge conflict\([Xie et al\. 2024](https://arxiv.org/html/2609.27590#bib.bib17);[Fu et al\. 2026](https://arxiv.org/html/2609.27590#bib.bib16)\), but forced\-choice behavior does not identify a parameter\-level mechanism\.
## 7Discussion
Near\-ceiling retrieval is a control outcome, not evidence that the diagnostic is uninformative\. MWE\-ECL intentionally makes the distant scenario easy to identify under an explicit question, then asks whether the same evidence changes a decision opposed by a familiar lexical default\. The English ceiling models show that the benchmark does not mechanically force an access–use gap; the lower OR∣\\midRet values of other panels provide the discriminating signal\. Likewise, perfect performance through 128K does not establish a model’s context limit: it only right\-censors joint ECL at the largest tested endpoint\.
Override answers a conditional question—whether context repairs a model’s own wrong default—and its denominator is consequently model\-specific\. That choice avoids treating a model with a different prior as if it faced the same conflicts, but it also prevents naive ranking by OR alone\. Fixed shared\-row and consensus\-hard analyses address denominator mismatch; preservation measures anchor\-induced corruption; and OR∣\\midRet removes matched rows with retrieval errors\. The agreement across these analyses supports the dissociation while leaving deployment\-level rankings descriptive\.
The broader implication is methodological\. Long\-context evaluations should pair an access control with a task in which retrieved evidence must alter an already plausible decision—for example, a distant state update paired with a locally plausible action whose correct choice the update reverses, plus a direct state query and a no\-state control that preserve the local decision surface\. This applies beyond MWEs to updated user preferences, policy exceptions, tool state, and document revisions\. Conditional correction and preservation then distinguish inaccessible state, accessible but behaviorally inert state, and harmful overwriting; these conditional metrics should accompany overall task accuracy, since accuracy alone cannot separate correction from preservation of an already\-correct default\. MWE\-ECL offers one controlled instantiation of this principle, not a claim that MWE errors directly predict agent failures or a complete measure of long\-context reasoning\.
## 8Limitations
The separate headline calls support a behavioral access–use dissociation but do not reveal the interpretation call’s hidden retrieval state; the three\-model same\-call control omits the prior head\. Forced choice measures decisions rather than mechanisms\. Moreover, the no\-anchor note and semantic anchor differ in content and length, so their contrast is a benchmark\-specific behavioral effect rather than a semantics\-only intervention on an internal prior\.
The diagnostic intentionally uses an idiom\-heavy set of strong\-default MWEs, neutral administrative filler, and only 25 English and 10 Chinese families\. Chinese difficulty is not matched to English, and the larger expansion lacks matched human validation\. Six headline panels are single\-pass API snapshots; coverage reflects API availability during the run window, and reasoning modes vary by endpoint\. Cross\-grid rankings are descriptive, and maximum\-endpoint ECL values are right\-censored tested\-range summaries rather than estimates of a monotone failure boundary\. Future work should broaden languages, document genres, response formats, and deployments, and test whether extended reasoning closes the override gap\.
## 9Conclusion
MWE\-ECL separates distant\-anchor recoverability, no\-anchor defaults, and anchor\-conditioned interpretation\. On prior\-conflict rows, near\-ceiling retrieval can coexist with substantially lower contextual override\. Targeted same\-call and semantic\-interference controls indicate that this pattern is not explained solely by invocation separation or simple anchor loss\. These findings are behavioral and deployment\-specific, but they support a broader evaluation principle: long\-context benchmarks should test not only whether distant evidence can be recovered, but whether it changes a competing local decision\.
## References
- Alibaba Cloud \(2026\)Alibaba CloudQwen3\.6\-Plus model information\.Note:Alibaba Cloud Model Studio DocumentationAccessed: 2026\-07\-28External Links:[Link](https://help.aliyun.com/zh/model-studio/qwen3-6-plus)Cited by:[§5\.1](https://arxiv.org/html/2609.27590#S5.SS1.p1.1)\.
- Almheiriet al\.\(2026\)S\. Almheiri, B\. Elbouardi, S\. Z\. Pranida, I\. Nikishina, A\. R\. B, P\. Krishnamurthy, M\. C\. Airlangga, R\. A\. Genadi, N\. P\. G\. Bao, A\. H\. Yari, H\. O\. Toyin, N\. Mukhituly, M\. Attia, B\. Hassan, A\. F\. Hidayatullah, T\. Kuribayashi, H\. Li, S\. Bhat, and F\. KotoMultilingual idioms in sentences and conversations across high\-, medium\-, and low\-resource languages\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 12363–12389\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.acl-long.564),[Link](https://aclanthology.org/2026.acl-long.564/)Cited by:[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px4.p1.1)\.
- Anet al\.\(2025\)C\. An, J\. Zhang, M\. Zhong, L\. Li, S\. Gong, Y\. Luo, J\. Xu, and L\. KongWhy does the effective context length of LLMs fall short?\.InThe Thirteenth International Conference on Learning Representations,External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2025/hash/884baf65392170763b27c914087bde01-Abstract-Conference.html)Cited by:[§1](https://arxiv.org/html/2609.27590#S1.p1.1)\.
- Baiet al\.\(2024\)Y\. Bai, X\. Lv, J\. Zhang, H\. Lyu, J\. Tang, Z\. Huang, Z\. Du, X\. Liu, A\. Zeng, L\. Hou, Y\. Dong, J\. Tang, and J\. LiLongBench: a bilingual, multitask benchmark for long context understanding\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 3119–3137\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.172),[Link](https://aclanthology.org/2024.acl-long.172/)Cited by:[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px1.p1.1)\.
- Baiet al\.\(2025\)Y\. Bai, S\. Tu, J\. Zhang, H\. Peng, X\. Wang, X\. Lv, S\. Cao, J\. Xu, L\. Hou, Y\. Dong, J\. Tang, and J\. LiLongBench v2: towards deeper understanding and reasoning on realistic long\-context multitasks\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 3639–3664\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.183),[Link](https://aclanthology.org/2025.acl-long.183/)Cited by:[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px1.p1.1)\.
- Baldwin and Kim \(2010\)T\. Baldwin and S\. N\. KimMultiword expressions\.InHandbook of Natural Language Processing, Second Edition,N\. Indurkhya and F\. J\. Damerau \(Eds\.\),pp\. 267–292\.External Links:[Document](https://dx.doi.org/10.1201/9781420085938-C12),[Link](https://doi.org/10.1201/9781420085938-C12)Cited by:[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px4.p1.1)\.
- Chenet al\.\(2026\)Z\. Chen, X\. Wu, J\. Jia, C\. Gao, Q\. Fu, D\. Zhang, and S\. HuLongBench Pro: a more realistic and comprehensive bilingual long\-context evaluation benchmark\.arXiv preprint arXiv:2601\.02872\.External Links:[Link](https://arxiv.org/abs/2601.02872)Cited by:[§1](https://arxiv.org/html/2609.27590#S1.p3.1),[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px2.p1.1)\.
- Constantet al\.\(2017\)M\. Constant, G\. Eryigit, J\. Monti, L\. van der Plas, C\. Ramisch, M\. Rosner, and A\. TodirascuMultiword expression processing: a survey\.Computational Linguistics43\(4\),pp\. 837–892\.External Links:[Document](https://dx.doi.org/10.1162/COLI%5Fa%5F00302),[Link](https://doi.org/10.1162/COLI_a_00302)Cited by:[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px4.p1.1)\.
- DeepSeek\-AI \(2026\)DeepSeek\-AIDeepSeek\-V4: towards highly efficient million\-token context intelligence\.External Links:2606\.19348,[Link](https://arxiv.org/abs/2606.19348)Cited by:[§5\.1](https://arxiv.org/html/2609.27590#S5.SS1.p1.1)\.
- Duet al\.\(2024\)K\. Du, V\. Snæbjarnarson, N\. Stoehr, J\. White, A\. Schein, and R\. CotterellContext versus prior knowledge in language models\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 13211–13235\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.714),[Link](https://aclanthology.org/2024.acl-long.714/)Cited by:[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px2.p1.1)\.
- Duet al\.\(2025\)Y\. Du, M\. Tian, S\. Ronanki, S\. Rongali, S\. B\. Bodapati, A\. Galstyan, A\. Wells, R\. Schwartz, E\. A\. Huerta, and H\. PengContext length alone hurts LLM performance despite perfect retrieval\.InFindings of the Association for Computational Linguistics: EMNLP 2025,Suzhou, China,pp\. 23281–23298\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.findings-emnlp.1264),[Link](https://aclanthology.org/2025.findings-emnlp.1264/)Cited by:[§1](https://arxiv.org/html/2609.27590#S1.p3.1),[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px2.p1.1)\.
- Fuet al\.\(2026\)Y\. Fu, H\. S\. Shahgir, H\. Liu, X\. Tang, Q\. He, and Y\. DongHarnessing the unseen: the hidden influence of intrinsic knowledge in long\-context language models\.Proceedings of the AAAI Conference on Artificial Intelligence40\(36\),pp\. 30762–30770\.External Links:[Document](https://dx.doi.org/10.1609/aaai.v40i36.40333),[Link](https://ojs.aaai.org/index.php/AAAI/article/view/40333)Cited by:[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px2.p1.1),[§6\.4](https://arxiv.org/html/2609.27590#S6.SS4.p7.1)\.
- Gemma Team \(2026\)Gemma TeamGemma 4 technical report\.External Links:2607\.02770,[Link](https://arxiv.org/abs/2607.02770)Cited by:[§5\.1](https://arxiv.org/html/2609.27590#S5.SS1.p1.1)\.
- Goldmanet al\.\(2024\)O\. Goldman, A\. Jacovi, A\. Slobodkin, A\. Maimon, I\. Dagan, and R\. TsarfatyIs it really long context if all you need is retrieval? towards genuinely difficult long context NLP\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,Miami, Florida, USA,pp\. 16576–16586\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.emnlp-main.924),[Link](https://aclanthology.org/2024.emnlp-main.924/)Cited by:[§1](https://arxiv.org/html/2609.27590#S1.p3.1),[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px2.p1.1)\.
- Haagsmaet al\.\(2020\)H\. Haagsma, J\. Bos, and M\. NissimMAGPIE: a large corpus of potentially idiomatic expressions\.InProceedings of the Twelfth Language Resources and Evaluation Conference,Marseille, France,pp\. 279–287\.External Links:[Link](https://aclanthology.org/2020.lrec-1.35)Cited by:[§1](https://arxiv.org/html/2609.27590#S1.p3.1),[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px4.p1.1),[§4](https://arxiv.org/html/2609.27590#S4.p1.1)\.
- Hagströmet al\.\(2025\)L\. Hagström, S\. V\. Marjanovic, H\. Yu, A\. Arora, C\. Lioma, M\. Maistro, P\. Atanasova, and I\. AugensteinA reality check on context utilisation for retrieval\-augmented generation\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Vienna, Austria,pp\. 19691–19730\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.968),[Link](https://aclanthology.org/2025.acl-long.968/)Cited by:[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px3.p1.1)\.
- Heet al\.\(2025\)W\. He, T\. K\. Vieira, M\. Garcia, C\. Scarton, M\. Idiart, and A\. VillavicencioInvestigating idiomaticity in word representations\.Computational Linguistics51\(2\),pp\. 505–555\.External Links:[Document](https://dx.doi.org/10.1162/coli%5Fa%5F00546),[Link](https://aclanthology.org/2025.cl-2.4/)Cited by:[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px4.p1.1)\.
- He \(2026\)W\. HeMore than meets the eye: measuring the semiotic gap in Vision\-Language Models via semantic anchorage\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),San Diego, California, United States,pp\. 15753–15767\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.acl-long.717),[Link](https://aclanthology.org/2026.acl-long.717/)Cited by:[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px4.p1.1)\.
- Hsiehet al\.\(2024\)C\. Hsieh, S\. Sun, S\. Kriman, S\. Acharya, D\. Rekesh, F\. Jia, and B\. GinsburgRULER: what’s the real context size of your long\-context language models?\.InFirst Conference on Language Modeling,External Links:[Link](https://openreview.net/forum?id=kIoBbc76Sy)Cited by:[§1](https://arxiv.org/html/2609.27590#S1.p1.1),[§1](https://arxiv.org/html/2609.27590#S1.p3.1),[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px1.p1.1)\.
- Liet al\.\(2025\)M\. Li, S\. Zhang, T\. Zhang, H\. Duan, Y\. Liu, and K\. ChenNeedleBench: evaluating LLM retrieval and reasoning across varying information densities\.Transactions on Machine Learning Research\.External Links:[Link](https://openreview.net/forum?id=cEvmIKsRw0)Cited by:[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px1.p1.1)\.
- Liet al\.\(2026\)Z\. Li, Y\. Tian, T\. Lan, J\. Li, Z\. Duo, G\. Gao, and X\. SuBorn pragmatic, trained to hallucinate? quantifying the origins of contextual bias in LLMs via the PaCE benchmark\.InFindings of the Association for Computational Linguistics: ACL 2026,pp\. 19211–19235\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.findings-acl.959),[Link](https://aclanthology.org/2026.findings-acl.959/)Cited by:[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px4.p1.1)\.
- Liuet al\.\(2024\)N\. F\. Liu, K\. Lin, J\. Hewitt, A\. Paranjape, M\. Bevilacqua, F\. Petroni, and P\. LiangLost in the middle: how language models use long contexts\.Transactions of the Association for Computational Linguistics12,pp\. 157–173\.External Links:[Document](https://dx.doi.org/10.1162/tacl%5Fa%5F00638),[Link](https://aclanthology.org/2024.tacl-1.9/)Cited by:[§1](https://arxiv.org/html/2609.27590#S1.p3.1),[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px1.p1.1)\.
- Miet al\.\(2025\)M\. Mi, A\. Villavicencio, and N\. S\. MoosaviRolling the DICE on idiomaticity: how LLMs fail to grasp context\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 7314–7332\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.362),[Link](https://aclanthology.org/2025.acl-long.362)Cited by:[§1](https://arxiv.org/html/2609.27590#S1.p3.1),[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px4.p1.1),[§4](https://arxiv.org/html/2609.27590#S4.p1.1)\.
- MiniMax \(2026\)MiniMaxMiniMax M3: frontier coding, 1m context, native multimodality\.Note:MiniMax ResearchAccessed: 2026\-07\-28External Links:[Link](https://www.minimax.io/blog/minimax-m3)Cited by:[§5\.1](https://arxiv.org/html/2609.27590#S5.SS1.p1.1)\.
- Modarressiet al\.\(2025\)A\. Modarressi, H\. Deilamsalehy, F\. Dernoncourt, T\. Bui, R\. A\. Rossi, S\. Yoon, and H\. SchuetzeNoLiMa: long\-context evaluation beyond literal matching\.InProceedings of the 42nd International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.267,pp\. 44554–44570\.External Links:[Link](https://proceedings.mlr.press/v267/modarressi25a.html)Cited by:[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px1.p1.1)\.
- Qwen Team \(2026\)Qwen TeamQwen3\.6\-35B\-A3B: agentic coding power, now open to all\.External Links:[Link](https://qwen.ai/blog?id=qwen3.6-35b-a3b)Cited by:[§5\.1](https://arxiv.org/html/2609.27590#S5.SS1.p1.1)\.
- Saget al\.\(2002\)I\. A\. Sag, T\. Baldwin, F\. Bond, A\. Copestake, and D\. FlickingerMultiword expressions: a pain in the neck for nlp\.InComputational Linguistics and Intelligent Text Processing,Lecture Notes in Computer Science, Vol\.2276,pp\. 1–15\.External Links:[Document](https://dx.doi.org/10.1007/3-540-45715-1%5F1),[Link](https://doi.org/10.1007/3-540-45715-1_1)Cited by:[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px4.p1.1)\.
- Sunet al\.\(2026\)K\. Sun, F\. Bai, and M\. DredzeTask matters: knowledge requirements shape LLM responses to context–memory conflict\.InFindings of the Association for Computational Linguistics: ACL 2026,pp\. 4154–4176\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.findings-acl.202),[Link](https://aclanthology.org/2026.findings-acl.202/)Cited by:[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px2.p1.1)\.
- Torunoğlu\-Selametet al\.\(2026\)D\. Torunoğlu\-Selamet, D\. Arslan, R\. Wilkens, W\. He, D\. Eryiğit, T\. Pickard, A\. S\. Pagano, A\. Villavicencio, G\. Eryiğit,et al\.A parallel cross\-lingual benchmark for multimodal idiomaticity understanding\.InProceedings of the Fifteenth Language Resources and Evaluation Conference \(LREC 2026\),pp\. 9434–9448\.External Links:[Document](https://dx.doi.org/10.63317/5cvnbcoktfo2),[Link](https://doi.org/10.63317/5cvnbcoktfo2)Cited by:[§4](https://arxiv.org/html/2609.27590#S4.p3.1)\.
- Wuet al\.\(2025\)D\. Wu, H\. Wang, W\. Yu, Y\. Zhang, K\. Chang, and D\. YuLongMemEval: benchmarking chat assistants on long\-term interactive memory\.InThe Thirteenth International Conference on Learning Representations,External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2025/hash/d813d324dbf0598bbdc9c8e79740ed01-Abstract-Conference.html)Cited by:[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px3.p1.1)\.
- Xiaoet al\.\(2025\)G\. Xiao, J\. Tang, J\. Zuo, J\. Guo, S\. Yang, H\. Tang, Y\. Fu, and S\. HanDuoAttention: efficient long\-context LLM inference with retrieval and streaming heads\.InThe Thirteenth International Conference on Learning Representations,External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2025/hash/5c1ddd2e59df46fd2aa85c833b1b36ed-Abstract-Conference.html)Cited by:[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px1.p1.1)\.
- Xieet al\.\(2024\)J\. Xie, K\. Zhang, J\. Chen, R\. Lou, and Y\. SuAdaptive chameleon or stubborn sloth: revealing the behavior of large language models in knowledge conflicts\.InThe Twelfth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=auKAUJZMO6)Cited by:[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px2.p1.1),[§6\.4](https://arxiv.org/html/2609.27590#S6.SS4.p7.1)\.
- Yang \(2024\)Z\. YangRetrieval or holistic understanding? DOLCE: differentiate our long context evaluation tasks\.arXiv preprint arXiv:2409\.06338\.Cited by:[§1](https://arxiv.org/html/2609.27590#S1.p3.1),[§2](https://arxiv.org/html/2609.27590#S2.SS0.SSS0.Px2.p1.1)\.
- Zhipu AI \(2026a\)Zhipu AIGLM\-4\.7\.Note:Zhipu AI Open Platform DocumentationAccessed: 2026\-07\-28External Links:[Link](https://docs.bigmodel.cn/cn/guide/models/text/glm-4.7)Cited by:[§5\.1](https://arxiv.org/html/2609.27590#S5.SS1.p1.1)\.
- Zhipu AI \(2026b\)Zhipu AIGLM\-5\.1\.Note:Zhipu AI Open Platform DocumentationAccessed: 2026\-07\-28External Links:[Link](https://docs.bigmodel.cn/cn/guide/models/text/glm-5.1)Cited by:[§5\.1](https://arxiv.org/html/2609.27590#S5.SS1.p1.1)\.
## Appendix AScope and Reading Guide
This document supplies the evidence needed to audit the paper’s behavioral claims without reproducing internal run logs\. It focuses on five questions: \(1\) whether retrieval and override differ on identical conflict rows, \(2\) whether the result depends on model\-specific conflict sets, \(3\) whether same\-call and hard\-cue controls preserve the pattern, \(4\) whether repeated runs and the expanded catalog support robustness, and \(5\) whether item labels and prompt construction pass independent checks\. The separate research code\-and\-data archive contains canonical row\-level outputs, analysis scripts, checksums, and a paper\-to\-file map\. That archive is not included in this arXiv submission\.
The primary unit is a matched case–distance identity\. Accuracy uses all valid task\-specific rows; override conditions on cases where the no\-anchor prior is wrong; OR∣\\midRet additionally requires correct matched retrieval\. Confidence intervals use 10,000 whole\-family bootstrap resamples\. Exact within\-model tests apply sign flips to family\-level paired differences and use Holm correction over the eight prespecified headline panels\.
## Appendix BEvaluation Snapshot
All prompts request one A/B label with temperature zero and a four\-token visible answer budget\. The English headline grid contains 50 cases, 25 MWE families, three task heads, and seven distances from 0 to 128K \(1,050 records per model\)\. The API\-interface panel is a time\-stamped deployment comparison rather than an architecture ranking\. Exact request records and completed outputs are in the code\-and\-data package, together with endpoint identifiers and access dates; superseded retries and cost logs are excluded\. Separate prompt\-fit, hard\-cue, expansion, and Chinese regimes are not pooled with this grid\.
Table S1:Headline English deployment snapshots\. Configuration\-mismatched stress regimes are reported separately and are not pooled with this grid\.
## Appendix CPrimary Statistical Accounting
Table[S2](https://arxiv.org/html/2609.27590#A3.T2)gives the interval accounting omitted from the compact main table\. Every joint\-ECL entry passes at the largest tested endpoint underτ=0\.8\\tau=0\.8and is therefore a right\-censored tested\-range summary, not a monotone failure\-boundary estimate\.
The canonical headline file contains exactly 8,400 rows: each of the eight panels contributes 50 cases, three task heads, and seven distances\. All planned rows returned a parseable A/B label; there are no request errors or post\-hoc row exclusions in this grid\. Option order is counterbalanced within each case–distance cell, and all three heads are joined by the same case, distance, and option\-order identity before a conditional metric is computed\.
Override uses a model\-specific denominator because a conflict row is defined by that model’s own no\-anchor prediction\. Table[S4](https://arxiv.org/html/2609.27590#A3.T4)therefore reports both pairwise conflict\-set overlap and a fixed 182\-identity consensus\-hard subset\. The former shows how similarly models expose their local defaults; the latter supplies a common\-denominator interpretation check\. Neither substitutes for override: they are cross\-model sensitivity analyses for a metric whose primary scientific target is within\-model correction\.
Table S2:Full English headline\-grid accounting\. Intervals use whole\-family bootstrap resampling; override is computed on each model’s prior\-conflict set\.Table S3:Within\-model retrieval\-control versus override tests on identical prior\-conflict rows\. Discordances count matched retrieval\-correct/interpretation\- wrong and interpretation\-correct/retrieval\-wrong events\.Table S4:Conflict\-set accounting on the shared 0–128K grid\. Consensus accuracy is fixed\-subset interpretation accuracy, not a model\-conditioned override rate\. Mean Jaccard summarizes pairwise conflict\-set overlap\.
Table S5:Prior\-to\-interpretation transitions \(left\) and preservation with family\-macro sensitivity \(right\)\. W/C denote wrong/correct no\-anchor priors\. Preservation conditions on prior\-correct rows; corruption is its complement\.The transition counts make the headline pattern concrete\. Across the five non\-ceiling panels, prior\-conflict correction failures range from 11 to 34, whereas prior\-correct corruption ranges from zero to four\. Interpretation loss is therefore concentrated on rows that require changing an initially wrong default, rather than on indiscriminate degradation after an anchor is added\. Family\-macro override closely tracks the micro\-pooled estimate, indicating that the result is not driven by one unusually frequent MWE family\.
No\-anchor stability at 128K
Interpretation retention at 128K
Exact matched\-grid rates by distance \(Interpretation/Retrieval/Override\)
Table S6:Sense and endpoint diagnostics\. Top: literal–figurative analysis; subscripts L/F denote paired variants within each family, and Holm\-adjustedppvalues come from exact two\-sided sign\-flip tests on family\-level interpretation differences\. Bottom left: no\-anchor semantic agreement with 0K, disagreement \(flip\), semantic\-figurative rate, and raw option\-A rate\. Bottom right: interpretation retention among items correct at 0K and on matched 128K retrieval; failure is one minus retention\. Bottom: exact rates underlying the distance curves, where override uses each model–distance cell’s prior\-conflict denominator\. Undefined family\-level conflict denominators are omitted only from the corresponding OR∣\\midRet statistic\.At 128K, no\-anchor semantic choices agree with their 0K counterparts on 0\.820–0\.960 of cases, while raw option\-A rates remain between 0\.260 and 0\.400\. This separates semantic default movement from a fixed answer\-position bias\. The retention analysis asks a different question: among items already correct at 0K and still retrieval\-correct at 128K, how often does interpretation remain correct? Retention ranges from 0\.872 to 1\.000, so endpoint failures are not solely inherited from items that were already wrong at short context\.
## Appendix DSame\-Call and Semantic\-Interference Controls
The same\-call control places retrieval and interpretation decisions after one shared document and counterbalances query order\. It removes invocation separation but does not expose the prior head\. At 64K and 128K, retrieval is perfect in all DeepSeek V4 Pro and GLM\-5\.1 cells and in three of four MiniMax\-M3 cells; interpretation remains below retrieval for DeepSeek V4 Pro and MiniMax\-M3 under both query orders\. The hard\-cue control instead changes semantic interference while retaining a matched neutral condition\.
The same\-call grid contains 600 completed responses: three models, two distances, two query orders, and 50 cases per cell\. Both decisions are decoded from one response, so the comparison cannot be attributed to separate model invocations\. Query order is counterbalanced because producing one label can itself condition the second; the order\-specific entries in Table[S7](https://arxiv.org/html/2609.27590#A4.T7)expose rather than average away this possibility\.
Table S7:Same\-call control; entries are interpretation/retrieval accuracy over 50 English cases per cell\.Hard\-cue records replace the neutral surroundings with either a target cue that supports the anchor\-licensed sense or a foil cue that reinforces the competing local prior\. Deltas are paired by family, case, distance, and task and are always condition minus neutral\. Thus a negative foil\-interpretation delta denotes greater interference, while a negative foil\-prior delta means the foil moved the no\-anchor default in the same direction\. Whole\-family resampling keeps all observations from an MWE family together\.
ModelFamiliesPairednnΔ\\Deltatarget Int\. \[95% CI\]Δ\\Deltafoil Int\. \[95% CI\]Δ\\Deltafoil Prior \[95% CI\]Δ\\Deltafoil Ret\. \[95% CI\]Matched 10\-family pilotDeepSeek V4 Flash10600\.067 \[0\.000, 0\.167\]\-0\.100 \[\-0\.217, 0\.000\]\-0\.350 \[\-0\.500, \-0\.200\]0\.000 \[0\.000, 0\.000\]GLM\-5\.110600\.033 \[0\.000, 0\.083\]0\.033 \[0\.000, 0\.083\]\-0\.167 \[\-0\.317, \-0\.050\]0\.000 \[0\.000, 0\.000\]25\-family cue\-triad runsDeepSeek V4 Pro251000\.060 \[0\.020, 0\.100\]\-0\.080 \[\-0\.140, \-0\.020\]\-0\.250 \[\-0\.320, \-0\.180\]\-0\.140 \[\-0\.200, \-0\.080\]Qwen3\.6\-Plus251000\.000 \[\-0\.030, 0\.030\]\-0\.060 \[\-0\.130, \-0\.010\]\-0\.380 \[\-0\.460, \-0\.300\]0\.000 \[0\.000, 0\.000\]GLM\-4\.7251000\.000 \[0\.000, 0\.000\]0\.000 \[0\.000, 0\.000\]0\.020 \[\-0\.060, 0\.090\]0\.000 \[0\.000, 0\.000\]MiniMax\-M3251000\.050 \[\-0\.010, 0\.110\]0\.000 \[\-0\.050, 0\.050\]\-0\.110 \[\-0\.200, \-0\.040\]\-0\.060 \[\-0\.110, \-0\.020\]Qwen3\.6\-35B251000\.050 \[0\.000, 0\.110\]0\.020 \[\-0\.020, 0\.060\]\-0\.330 \[\-0\.410, \-0\.250\]0\.000 \[0\.000, 0\.000\]Gemma\-4\-26B251000\.040 \[\-0\.010, 0\.100\]\-0\.010 \[\-0\.050, 0\.030\]\-0\.190 \[\-0\.280, \-0\.100\]0\.000 \[0\.000, 0\.000\]
Table S8:Paired hard\-cue changes relative to separately executed neutral controls\. Intervals use 10,000 whole\-family bootstrap resamples\. The pilot and all\-family panels differ in scope and are not pooled into a ranking\.The paired cue deltas delimit two competing explanations\. Target cues do not systematically harm interpretation and improve it for DeepSeek V4 Pro in the all\-family run\. Foil cues, by contrast, reduce interpretation for DeepSeek V4 Pro and Qwen3\.6\-Plus, while changing the no\-anchor prior more strongly for most panels\. Retrieval remains unchanged for four of six all\-family panels and drops for DeepSeek V4 Pro and MiniMax\-M3\. Thus semantic interference can affect both recoverability and decision use, but the two effects are not identical\. Because the pilot and all\-family runs differ in catalog coverage and execution window, the table supports within\-panel cue contrasts rather than a model ranking\.
## Appendix ERepeatability and Position Pilot
Three\-pass endpoint checks quantify API\-interface nondeterminism for the two open\-weight model endpoints\. The 128K passes reproduce identical aggregate accuracies and labels\. The position probe is deliberately descriptive because each cell contains only 20 matched items\.
Each repeatability cell consists of three independently executed 50\-record passes\. Agreement is the fraction of item identities receiving the same label in all passes, whereas the reported standard deviation summarizes pass\-level accuracy\. The small 64K Qwen variation therefore reflects a few label changes, not missing records; all four 128K interpretation/retrieval cells have perfect cross\-pass label agreement\. These checks quantify deployment variation and do not turn a three\-pass snapshot into a model\-level stability guarantee\.
Table S9:Endpoint repeated\-run accuracy and per\-item label agreement over three 50\-record passes per cell\. Qwen\-35B and Gemma\-26B abbreviate the model labels used in the main paper\.Table S10:Preliminary Qwen3\.6\-35B\-A3B anchor\-position probe\. The small\-NNdesign does not support a general position\-effect claim; Ret\. correct is a count out ofnn, while IUR and override are rates\.At 128K, all four interpretation/retrieval repeatability cells have identical labels across the three passes\. The 64K Qwen cells show the only pass\-level variation, with agreement of 0\.920–0\.980 depending on the head\. The position pilot contains no monotone late\-to\-early ordering: at 8K, override ranges from 0\.70 to 0\.90 across positions, while at 64K it is 0\.60–0\.70\. These 20\-item cells are useful as a failure\-mode check but are too small to estimate a general positional curve\.
## Appendix FPrompt\-Fit Stress Beyond 128K
Prompt\-fit runs test whether the access–use pattern remains visible when a deployment is evaluated near its practical serving limit\. They use the same three task heads and 50 English cases, but endpoint\-specific filler sizes and tokenizers make them descriptive extensions rather than additional matched\-grid distance cells\. All 600 responses complete without provider errors or unparsed labels\.
Table S11:Prompt\-fit stress outside the matched grid\. Requested filler is 217,088 tokens for the 256K rows, 512K for the 512K row, and 960K for the 1M row\. Each row contains 150 records \(50 cases across three heads\), with no request errors or unparsed labels\. Exact physical prompt\-token ranges are retained for reproducibility\.All four prompt\-fit rows retain perfect explicit retrieval, whereas interpretation ranges from 0\.860 to 0\.940\. The DeepSeek V4 Flash 1M row is not worse than its 512K row, reinforcing that these independently prompted cells should not be read as a monotone degradation curve\. Their role is narrower: the access–use dissociation remains observable near endpoint\-specific serving limits when the complete prompt physically fits\.
## Appendix GDataset Validation and Catalog Expansion
The English primary catalog is fixed before model evaluation\. Two independent annotators selected the anchor\-licensed answer for all 50 items without seeing model outputs; both matched all intended labels and raised no flags\. Automated checks verify schema, task/distance balance, option order, and prompt structure\. A static audit finds no verbatim answer\-option leakage and no target\-MWE leakage from the neutral filler bank\. Four literal anchors repeat the physical target expression by design and are reported in the released audit\.
AuditEnglishChineseCases reviewed5020Independent annotators22Judgments matching intended labels100/10040/40Flagged items00Verbatim option leakage00Target\-MWE leakage from neutral filler00Table S12:Human label validation and static leakage checks\. Annotators were blind to model outputs; the two languages were reviewed separately\.The audited expansion adds 60 English cases from 30 additional families and is evaluated separately at 0/8K/64K\. It broadens lexical and constructional coverage but is not difficulty\-matched to the primary suite and is not used for headline rankings\.
Table S13:Audited 60\-case catalog\-expansion reproduction\. Each model has 540 records, zero request errors, and zero unparsed responses\. Primary OR is shown only for context; the two catalogs are not pooled\.A descriptive inventory\-sensitivity split is included in the released derived tables\. Because the inventory\-absent group is not a held\-out set, it does not support a contamination claim and is not used in the paper’s conclusions\.
Table S14:Aggregate results on the separately reported 20\-case, 10\-family Chinese subset\. The subset uses the same three heads and seven nominal distances, but is neither a translation nor a difficulty\-matched counterpart of the English catalog\.The Chinese subset reproduces the qualitative contrast for several panels: DeepSeek V4 Pro, MiniMax\-M3, Qwen3\.6\-35B, and Gemma\-4\-26B all have lower override than retrieval, while three ceiling panels remain at 1\.000\. DeepSeek V4 Flash is near ceiling on both measures\. This is a multilingual robustness check, not evidence that the two catalogs have equal difficulty: the Chinese items are independently sourced, smaller, and reported only in aggregate\.
## Appendix HPrompt and Release Boundary
The system instruction is: “Use the entire document, including distant context, to choose the correct interpretation\. Reply with exactly one letter: A or B\. Do not explain\.” The user prompt contains an anchor note, deterministic neutral filler, a final event, the task question, and two options\. Retrieval replaces the interpretation question with a direct anchor\-scenario query; prior\-control replaces the anchor with a note stating that no disambiguating background is provided\. Complete prompt records are included in the data package rather than duplicated here\.
The Chinese subset is reported only in aggregate because its source inventory uses controlled release\. It contains 20 cases across 10 families, is not a difficulty\-matched translation of the English suite, and is analyzed separately\. Two native\-Chinese annotators matched all intended labels with no flags\. The code\-and\-data package includes the transformation and validation code, but not the controlled inventory or row\-level Chinese outputs\.
## Appendix IInterpretive Boundary of the Controls
The controls address distinct alternatives and should not be combined into a single mechanistic claim\. Same\-call evaluation rules out invocation separation for its three tested deployments, but does not reveal the latent state used to produce either answer\. Hard\-cue deltas establish sensitivity to structured semantic interference, but a retrieval decrease in some panels prevents a pure integration\-only interpretation\. Repeated runs constrain deployment variation at selected endpoints, whereas the position pilot and prompt\-fit rows remain descriptive because their grids are deliberately smaller or configuration specific\.
Taken together, these checks support a behavioral conclusion: an explicit matched retrieval answer can be correct while the associated interpretation still follows the competing local default\. They do not identify a neural mechanism, prove that every interpretation error contains a recoverable hidden anchor representation, or establish a universal context\-length ranking\.
## Appendix JResidual Limitations
The benchmark is an idiom\-heavy forced\-choice diagnostic, not an estimate of average language understanding\. Conflict sets are model\-specific, deployment snapshots can drift, Chinese and English difficulty are not matched, and the expansion is not held out\. Separate\-call retrieval controls do not reveal the interpretation call’s hidden state; the same\-call control addresses invocation separation for three models but omits the prior head\. These constraints bound the claims to behavioral access and use under the reported deployments\.相似文章
长上下文LLM中的位置失败:推理基准测试的盲点
本论文识别出长上下文LLM推理基准测试中的一个盲点:它们未能控制任务在上下文中的位置,导致位置失败未被检测到。作者提出上下文旋转评估(CRE)来系统地改变任务位置、填充内容和上下文长度,揭示出当推理任务放置在长上下文中时,某些模型的准确率会严重下降。
坍塌后的知识:上下文与措辞塑造语言模型中事实知识的可访问性
该论文提出了知识回答分歧(Knowledge Answer Divergence,KAD),这是一种以事件为条件的评估协议,从语言模型采纳错误纠正的时刻分出探测分支,在八个本地模型检查点和六个托管端点上衡量错误的持续、恢复和弃答情况。研究发现,对话历史和措辞具有显著影响:中性历史分支大幅提高了 Llama 3.2 Instruct 3B 的回答正确率,而提示线索的位置使 Qwen 3.5 4B 采纳错误纠正的比例变化了 27.3 个百分点。
上下文归因如何处理模型已知的知识
本文介绍了一个评估协议,包含四个新指标和一个基准数据集,用于评估LLM的上下文归因方法,并表明当上下文与训练数据重叠时这些方法会失败。
研究求助:观察——语义密集上下文无需越狱提示即可引发强后期层发散 [D]
一项实证研究表明,长篇幅、语义密集且良性的文本能够偏移模型的潜在空间并绕过对齐,促使其生成原本被屏蔽的评论。作者并非专家,请求对其指标进行审计,以区分真正的语义劫持与假象。
论词汇性在大语言模型中的持续影响
本文研究了词汇重叠(而非语义内容)如何影响跨层和跨架构的大语言模型表示,并证明即使在为语义相似性训练的模型中,这种词汇效应依然存在,导致下游任务性能下降。