{"study": "Agent World Model in the Wild: Computer-Use Agent Environment Synthesis", "link": "https://arxiv.org/abs/2610.26828v1", "classification": "Indexed", "indexed": "2026-10-25", "receipt": ["Read: abstract, author list and released-file list, version v1", "Released output files listed: 56", "Not recounted"]}
{"study": "Behind Aggregate Scores: What Systematically Varying Arena-Hard's LLM Judge Reveals", "link": "https://arxiv.org/abs/2610.30512v1", "classification": "Indexed", "indexed": "2026-10-25", "receipt": ["Read: full text, version v1, and their released files (34 output files)", "Printed numbers against their released files: CLEAN: 17,000/17,000 verdicts parse identically. Table 1 rebuilds to 2 dp. 21.2%, +28.2 pp, +6.7 pp, -4.09 points", "Measured on our setup: fingerprint carrier, runs 1-2: same backend 17.5% vs changed 27.3% outcome change, +9.8 pp [+2.1, +17.6] ABOVE. Margin-stratified (defined after the data) +8.0 pp [+0.7, +15.3] ABOVE. Runs 1-3 / 2-3 CANNOT_TELL", "Data: /iris/runs/recounts.zip"]}
{"study": "Can LLMs predict the outcomes of machine learning experiments?", "link": "https://arxiv.org/abs/2610.22599v1", "classification": "Indexed", "indexed": "2026-10-25", "receipt": ["Read: abstract, author list and released-file list, version v1", "Released output files listed: 1", "Not recounted"]}
{"study": "Can LLMs Tell Stories Out of Order? A Corpus and Benchmark for Nonlinear Narrative Generation", "link": "https://arxiv.org/abs/2610.21468v1", "classification": "Indexed", "indexed": "2026-10-25", "receipt": ["Read: full text, version v1, and their released files (57 output files)", "Printed numbers against their released files: CLEAN: judge tallies 31-14-11 and 23-22-11. Reader G1 share 72.1% of decisive", "Measured on our setup: item-level judge vs reader majority: G1 vs G2 0.657 [0.486, 0.800] CONSISTENT with chance, BELOW reader retest 0.930. G1 vs G0 0.839 [0.710, 0.968] ABOVE chance", "Data: /iris/runs/recounts.zip"]}
{"study": "CulturalRiddles: A Multicultural Benchmark of Riddles", "link": "https://arxiv.org/abs/2610.29968v1", "classification": "Indexed", "indexed": "2026-10-25", "receipt": ["Read: abstract, author list and released-file list, version v1", "Released output files listed: 48", "Not recounted"]}
{"study": "GlobeBench: Evaluating Text-Based World Model Capabilities for Scalable Environment Simulations", "link": "https://arxiv.org/abs/2610.22708v1", "classification": "Indexed", "indexed": "2026-10-25", "receipt": ["Read: abstract, author list and released-file list, version v1", "Released output files listed: 117", "Not recounted"]}
{"study": "IndoNutriQA: An Indonesian Clinical Nutrition Question Answering Benchmark", "link": "https://arxiv.org/abs/2610.31256v1", "classification": "Indexed", "indexed": "2026-10-25", "receipt": ["Read: full text, version v1, and their released files (5,037 output files)", "Printed numbers against their released files: CLEAN within rounding: EM 22/22. F1 18/22 to the digit, 4 off by 0.01-0.02", "Measured on our setup: 5-shot gain ABOVE for all 8 LLMs, paired. The paper reads GPT-4o as matched by Llama-3.1-8B-Instruct under five-shot prompting because the confidence intervals overlap. Paired over the same items, Llama leads by +5.24 [+0.91, +9.81] ABOVE. Post-hoc labelled: Llama vs MedGemma CONSISTENT, MedGemma vs GPT-4o ABOVE", "Data: /iris/runs/recounts.zip"]}
{"study": "Jev-ToM: Evaluating Theory of Mind in Typed Decision Models", "link": "https://arxiv.org/abs/2610.18610v1", "classification": "Indexed", "indexed": "2026-10-25", "receipt": ["Read: full text, version v1, and their released files (25 output files)", "Printed numbers against their released files: CLEAN: overall accuracy for all 12 models rebuilds from per-question predictions", "Measured on our setup: Jev lead +1.59 [+0.25, +2.89] over GPT-3.5 and +2.36 [+0.98, +3.75] over GPT-5.6-Luna, ABOVE. The paper says \"Removing native controls reverses Jev’s lead over GPT-3.5, leaving 69.1% against 70.9%.\" This rebuilds when the 920 control items are identified from the released category field: 69.08 vs 70.87, difference -1.79 [-3.54, -0.19] by family_id. Our earlier CANNOT_TELL is withdrawn.", "Data: /iris/runs/recounts.zip"]}
{"study": "Legal Guidance in the Wild: How Users Seek Legal Help in Real-World LLM Conversations", "link": "https://arxiv.org/abs/2610.25159v1", "classification": "Indexed", "indexed": "2026-10-25", "receipt": ["Read: full text, version v1, and their released files (1 output file)", "Printed numbers against their released files: CLEAN: all guidance metrics rebuild from 200 per-item predictions", "Measured on our setup: silver GPT labeller scored as a system: 0.960 accuracy, +0.115 to +0.145 over every trained classifier (ABOVE), on a gold anchored to it. Anchoring share CANNOT_TELL. Corrected-item test n=8 CANNOT_TELL", "Data: /iris/runs/recounts.zip"]}
{"study": "MapSteal: Map-based Autonomous Knowledge Stealing Attack Against Large Language Models", "link": "https://arxiv.org/abs/2610.25626v1", "classification": "Indexed", "indexed": "2026-10-25", "receipt": ["Read: full text, version v1, and their released files (182 output files)", "Printed numbers against their released files: CLEAN: Table 2 means/SDs and the 12-of-18 count rebuild from per-run accuracies. Item level CANNOT_TELL (keys not released)", "Measured on our setup: with seeds as units, MapSteal leads the best other method ABOVE on 2 of 18 settings, CONSISTENT on 15, BELOW on 1", "Data: /iris/runs/recounts.zip"]}
{"study": "Mixing Matters: Evidence Position Bias Across Sequence Mixers in Long-Context Question Answering", "link": "https://arxiv.org/abs/2610.31135v1", "classification": "Indexed", "indexed": "2026-10-25", "receipt": ["Read: full text, version v1, and their released files (3 output files)", "Printed numbers against their released files: CLEAN: Table 1 edges, floors, ceilings and paired contrasts rebuild from 229,700 records", "Measured on our setup: known-good: separate same-host runs identical on 8,000 of 8,000 question-position cells. Cross-GPU copy (A40 against A10G) changes 76 of 8,000 cells (0.95%), primacy gap +0.0019 [-0.0050, +0.0088] CONSISTENT, so it did not move the score. Family gap ABOVE on 2 of 5 prompt variants, CONSISTENT on 3. Prompt variants also changed GPU (T4/L40S): GPU share CANNOT_TELL", "Data: /iris/runs/recounts.zip"]}
{"study": "NepPlural: A Human-Anchored Benchmark for Persona Annotation in Nepali Migration Discourse", "link": "https://arxiv.org/abs/2610.15732v1", "classification": "Indexed", "indexed": "2026-10-25", "receipt": ["Read: full text, version v1, and their released files (97 output files)", "Printed numbers against their released files: CLEAN: Table 4 forced/abstain macro-F1, No-Persona F1 and false-abstention rates rebuild from per-comment outputs", "Measured on our setup: abstention gain +0.20 to +0.27 macro-F1 ABOVE for GPT, Gemini, Claude and majority. Gemini ahead of both ABOVE. GPT vs Claude CONSISTENT (+0.025 [-0.027, +0.078])", "Data: /iris/runs/recounts.zip"]}
{"study": "Online LLM Personalization from Pairwise Preferences Across Users", "link": "https://arxiv.org/abs/2610.17203v1", "classification": "Indexed", "indexed": "2026-10-25", "receipt": ["Read: full text, version v1, and their released files (16 output files)", "Printed numbers against their released files: author-author 69.6% CLEAN. Opus 73.7 vs 74.0 CLEAN. Sol rebuilds (74.5) only under the favorite rule, found after the data", "Measured on our setup: both chat judges vs author-author CONSISTENT (+3.3 [-2.8, +9.4], +4.0 [-1.9, +9.8]). The scoring rule flips the Sol/Opus order, neither resolves (-0.7 [-3.6, +2.1])", "Data: /iris/runs/recounts.zip"]}
{"study": "PRISM: Transfer and Limits of Activation Interventions in Tool-Calling Language Models", "link": "https://arxiv.org/abs/2610.20808v1", "classification": "Indexed", "indexed": "2026-10-25", "receipt": ["Read: abstract, author list and released-file list, version v1", "Released output files listed: 55", "Not recounted"]}
{"study": "SoK: PIAtlas: Decomposing the Unbounded Prompt Injection Attack Space", "link": "https://arxiv.org/abs/2610.24999v1", "classification": "Indexed", "indexed": "2026-10-25", "receipt": ["Read: full text, version v1, and their released files (52 output files)", "Printed numbers against their released files: CLEAN: all ten labeller calibration agreements rebuild from the 64-record runs", "Measured on our setup: batched labelling (4 or 8 per call) agrees with single-record labels 0.02-0.03 micro-F1 below the single-record run-to-run floor, BELOW on all four readings", "Data: /iris/runs/recounts.zip"]}
{"study": "Token-Set Choice Confounds POPE: A Systematic Audit of Yes/No Extraction in VLM Hallucination Evaluation", "link": "https://arxiv.org/abs/2610.25964v2", "classification": "Indexed", "indexed": "2026-10-25", "receipt": ["Read: full text, version v2, and their released files (27 output files)", "Printed numbers against their released files: CLEAN: every F1/precision/recall/yes-rate, T/G/rho/phi, token-id and zero-transition figure rebuilds from the per-question records", "Measured on our setup: shift 6.13 [4.43, 7.83] adversarial, 5.37 [3.70, 7.03] popular, 3.15 [1.50, 4.80] random, ABOVE. The paper says the shift \"exceeds the adversarial gain reported by five of the six papers we surveyed\". Against our interval four of six gains lie below the lower bound, one (+5.76) inside and one above, so the resolved count is four of six with one undecided. One 4-bit model, so CANNOT_TELL for a matched evaluation", "Data: /iris/runs/recounts.zip"]}
{"study": "Verdict: Encode Once, Decide Many", "link": "https://arxiv.org/abs/2610.20778v1", "classification": "Indexed", "indexed": "2026-10-25", "receipt": ["Read: abstract, author list and released-file list, version v1", "Released output files listed: 528", "Not recounted"]}