HyperFeed

Orchestra combines RegNetAgents and CASCADE into a multi-agent workflow for finding likely cancer regulators. RegNetAgents supplies gene-network topology, while CASCADE checks four independent experimental evidence sources. Candidates supported by at least two CASCADE sources showed stronger agreement with OncoKB cancer-gene status, especially in the STAD replication. The study argues that corroboration across agents can improve biological discovery over single-source ranking.

• BRCA and COAD focal genes reached odds ratio 2.89 with adjusted p=0.0166
• STAD replication strengthened the association to odds ratio 5.82
• CASCADE checks LINCS, DepMap, super-enhancer, and DoRothEA evidence
• Single-source evidence was not diagnostic for focal or negative-control genes

A new study tests deterministic checks before LLM agents execute shell commands or apply code edits. Its verifier catches 95.8% of invalid commands across 9,930 examples, while location-based edits often corrupt files without errors. The work argues that agents should ground actions before execution and abstain when checks cannot establish correctness.

• 9930 commands and 482 tools formed the shell-command evaluation set
• 95.8% invalid-command detection came with a 10.0% false-positive rate
• Syntax and binary checks caught half of errors with zero false positives
• Line-number edits corrupted 99.1% of files after a one-line shift
• Function-name edits targeted the wrong function in 12.7% of cases

ORQA introduces a source-traceable benchmark for testing occupation-level knowledge in large language models. It builds questions from trusted websites tied to O*NET occupations, then combines automated generation with human review. Claude Opus 4.6, GPT-5.4, and Claude Sonnet 4.6 lead the evaluation at roughly 58% to 62%. Results vary sharply by occupation, exposing gaps hidden by broad capability tests.

• 116 occupations span all 21 major SOC groups
• 480 questions draw from 187 occupation-specific websites
• Healthcare occupations reach 78% performance in the evaluation
• Office and Administrative Support occupations score about 40%
• Smaller open-weight models achieve approximately 33% to 41%

NDT Factory is a multi-agent system that builds executable network digital twins from semantic models. It targets autonomous network management by generating analysis logic for changing service intents instead of relying on fixed implementations. A Call Admission Control test produced deterministic decisions, with 99.3% agreement against a reference system across 300 simulated intents.

• 100% compilation and test pass rates across multiple synthesis runs
• 300 simulated Network Service Intents evaluated in the case study
• 90% admission rate recorded during Call Admission Control simulation
• All rejected requests received correct attribution from the generated twin

HORIZON is a hierarchical agent built for Lux AI Season 3’s partially observable, multi-agent matches. It tracks hidden game state and opponent style across episodes, combining memory, graph attention, exploration, and policy mixtures. The system separates immediate control from cross-match reasoning, helping it adapt to randomized dynamics and best-of-five competition. PPO training in a large JAX simulator produced stronger results than recurrent and feed-forward baselines.

• Lux AI Season 3 uses randomized episode-level dynamics and best-of-five matches
• HORIZON combines dual-memory belief tracking with relic-centric graph attention
• Information-gain exploration guides actions when opponents and game parameters remain hidden
• Training used PPO in a large-scale JAX simulator
• Reported gains covered match wins, episode wins, adaptation, and league rating

DementiaCare-Bench introduces a video benchmark for understanding behavioral and psychological symptoms of dementia and their triggers. The dataset contains 94 clips from 56 caregiver training videos, with 2,023 transcript-grounded questions across nine symptom categories. Tests of 12 video-language models show that overall scores hide weak temporal reasoning, with accuracy dropping 17 points when ordered visual context is required.

• 56 training videos produce 94 clips spanning nine BPSD categories
• 2,023 questions use clinical claims grounded in verbatim transcript spans
• 77.7% were designed to require ordered frames, but only 34.8% did
• Best models reached 85% overall across the benchmark
• A leading open model scored at chance on caregiver-response judgments

Meddies-PII presents one million synthetic clinical documents across 17 languages for training multilingual PII extraction systems. The dataset covers nine PII labels and uses attribute-conditioned prompts with 13 deterministic validation gates. Its BIOES token classifier reached a mean exact-match F1 of 0.827 across 15 external benchmarks, versus 0.658 for the strongest baseline. The authors plan to release the dataset, model, framework, and evaluation code.

• 17 languages covered, including nine personally identifiable information labels
• 13 deterministic gates validate document structure and annotation consistency
• 0.827 mean F1 achieved across 15 external benchmarks
• 0.658 mean F1 recorded by the strongest evaluated baseline
• One million synthetic clinical documents make up Meddies-PII-Dataset

A new study finds multimodal language models collapse ambiguity when describing images in a Dixit-style task. Unlike human players, models tend to produce literal, over-specified clues that leave little room for multiple interpretations. They also rarely use culturally situated references, even when prompted to use allusion and figurative language. The findings point to limits in models’ cultural and communicative flexibility.

• arXiv paper 2609.12575 introduces a rubric for measuring calibrated ambiguity
• Dixit provides the study’s task for comparing human and model-generated clues
• Models almost never referenced culturally situated knowledge in their clues
• Prompts explicitly asked models to use allusion and figurative language

RMDM is a masked diffusion language model that uses continuous text representations to coordinate parallel token generation. The method encodes text into a Gaussian-normalized latent space, then conditions token updates on that global semantic signal. This targets a core weakness of masked diffusion sampling, where independently updated tokens can conflict. Results show larger quality gains when generation uses only a few steps.

• RMDM uses a pretrained encoder to create continuous semantic representations
• An invertible transform maps representations to a Gaussian prior for sampling
• Global latent guidance coordinates multiple masked-token updates during generation
• The method focuses on aggressive few-step sampling regimes

R2VC is a modular fact-checking system that separates retrieval, reasoning, verification, and confidence estimation. Its 8B implementation improved FEVER accuracy by 13.74% over a baseline while supporting citations and abstention. External candidate verification and calibration drove the largest gains, but retrieval errors, especially evidence about the wrong entity, remain the main failure mode.

• FEVER accuracy falls to 76.24% when verifier-based candidate selection is removed
• Calibration removal nearly doubles the Brier score to 0.161
• Hybrid sparse and dense retrieval searches Wikipedia for supporting evidence
• An external NLI cross-encoder selects among structured verdict candidates
• Manual review of 250 errors identified wrong-entity retrieval as dominant

Researchers propose a weakly supervised pipeline for finding dataset references in forced displacement and FCV documents. A lightweight model proposes candidate mentions, then a frontier LLM validates them and fixes extraction boundaries. Fine-tuning on the refined labels supports large-scale extraction without a large hand-labeled corpus. On 1,706 benchmark passages, the model reached 74.1% precision and 70.5% recall.

• 1,706 passages cover research, humanitarian, and operational documents
• 74.1% mention-level precision across the full benchmark
• 70.5% mention-level recall across the full benchmark
• Frontier LLM review corrects candidate boundaries and rejects false mentions

A new paper models factual hallucination as both missing coverage and lossy memory compression. Its bound shows that even facts a model has seen can be recalled incorrectly when stored in too few bits. The framework separates errors from unobserved facts and distorted memories, then tests its predictions with simulations and fact-injection probes in language models. It also examines retrieval, abstention, and long-context organization.

• Two error sources are separated: compression distortion on observed facts and missing coverage
• The bound uses N queries, K answers, M training facts, and B memory bits
• The compression term depends on the inverse rate-distortion function for uniform K-ary data
• Experiments vary fact load and effective trainable memory in modern language models

Researchers introduced a scalable method for ranking consonants by their impact on word intelligibility. The approach silences one consonant in isolated words, then measures how often ASR systems misrecognize the result. Tests across English, Spanish, German, and Czech used MMS, Whisper, and Qwen3-ASR, linking higher disruption to greater lexical contrast and lower phoneme frequency.

• Four languages tested: English, Spanish, German, and Czech
• Three ASR architectures evaluated: MMS, Whisper, and Qwen3-ASR
• Metric named mask-induced misrecognition rate, or MMR
• Partial Spearman analysis linked phoneme frequency negatively with MMR
• Functional load correlated positively with consonant masking disruption

MedCollab introduces a multi-agent framework that coordinates specialist and examination agents for full-cycle clinical diagnosis. The system uses IBIS to tie each diagnostic claim to patient evidence and medical knowledge, then links hypotheses with hierarchical disease relations. A verifier-guided consensus step filters unsupported reasoning and resolves conflicts. Tests on ClinicalBench and MIMIC-IV report gains over leading LLMs and medical agent baselines.

• IBIS protocol grounds each diagnostic position in patient-specific evidence and medical knowledge
• HDRC links hypotheses through disease progression, complications, and comorbidities
• Verifier module scores evidence support, medical plausibility, and logical conflicts
• Evaluated on ClinicalBench and MIMIC-IV clinical datasets
• Reports gains in accuracy, evidence consistency, and reasoning quality

RIPPLE is a method for improving agents that synthesize executable workflows without changing their underlying model. It maps failures to policy segments, applies targeted edits, and tests candidates against the same starting policy. Promising edits are then replayed after accepted updates, filtering out changes whose effects interact badly downstream. The approach targets safer persistent adaptation for multi-step agent workflows.

• Flow-HO provides the synthetic evaluation setting for RIPPLE
• Failed trajectories determine which predefined policy segment receives an edit
• Candidate edits are first compared using the same iteration-start policy
• Replay after accepted updates exposes downstream effects and edit interactions

Researchers used 25,125 Google Maps reviews to measure how residents experience food access across 49 Raleigh grocery stores. Topic modeling and zero-shot classification separated availability, accessibility, affordability, accommodation, and acceptability. The results show that nearby stores can create different experiences, even under identical policies, giving planners a scalable complement to geographic access measures.

• 49 grocery stores across Raleigh supplied the review data
• 85.4% agreement matched manual coding of review topics
• Five dimensions included affordability, accommodation, and acceptability
• Nearby stores in the same chain showed different perceived barriers

Researchers propose context-augmented training to help language models answer multi-hop questions over knowledge graphs. The method adds supporting triples from the same source text to each target relation, giving models broader evidence during supervision. An adaptive repair loop then targets unresolved one-hop errors and quarantines noisy triples. Tests use disease-specific graphs for Gastroparesis and Diabetes.

• GraphMERT extracted disease knowledge graphs for Gastroparesis and Diabetes
• Two variants, KGModel and CGModel, receive separate supervised fine-tuning
• Repair examples target unresolved one-hop failures identified by an LLM judge
• Supporting triples come from the same source text chunk as each target triple

Researchers introduced an open framework that finds and structures social tipping point evidence at the passage level. It combines document segmentation, passage classification, rewriting, criteria scoring, and semantic retrieval in one workflow. The system uses several language models but is a document analysis pipeline, not an autonomous agent. Tests covered 163 GPT-4.1-labeled passages and 51 expert-reviewed passages.

• 71.4% accuracy came from the iteratively augmented RoBERTa detector
• 6.137 composite score put the DistilBERT splitter ahead of three alternatives
• LLaMA 3.2 3B scores passages against five published criteria
• Milvus provides semantic retrieval, with Streamlit and MinIO supporting deployment

HypoKG combines KEGG, Rhea, and UniProt into a biochemical knowledge graph for testing evidence-grounded biomedical hypothesis generation. Researchers evaluated 13,200 hypotheses from six LLMs across four information settings and found a tradeoff between persuasive ideas and mechanistic support. Full biological paths produced more evidence-consistent hypotheses, while endpoint-only prompts often scored higher overall. Shuffling intermediate steps sharply reduced grounding, suggesting models used path structure.

• 550 enzyme-to-disease paths formed the benchmark
• 13,200 hypotheses came from six LLMs
• Five expert criteria scored each hypothesis from 1 to 5
• Shuffled paths reduced evidence grounding by 0.793
• The grounding drop was statistically significant at p < 0.001

GraphProfiler is an auditable LLM profiler that infers sensitive attributes from user post histories and cites the evidence behind each prediction. It builds a source-linked personal knowledge graph, connecting concepts and relationships back to originating posts. The system reached 86.7% attack success on SynthPAI and cited support for more than 98% of predictions. That traceability could enable targeted privacy fixes instead of rewriting entire histories.

• 86.7% attack success on the eight-attribute SynthPAI benchmark
• 84.6% attack success on the PANDORA dataset
• Over 98% of predictions included cited supporting evidence
• Removing cited posts cut success more than removing equal random sets

ASCIL is a post-ASR inference layer designed to stop assistants from acting on speech that only resembles a wake word. It combines acoustic, linguistic, device-context, and user-history signals before response generation. The system also learns from hesitation, silence, cancellations, and repeated commands, updating its corrections online without manual labels during use.

• 3,667 interactions formed the proprietary evaluation dataset
• 14 acoustic and contextual conditions covered the reported evaluation
• Human annotators supplied intentional and unintentional reference labels
• Online updates use inferred behavioral signals instead of manual annotations

EAR introduces an entity-aware way to partition documents for retrieval-augmented generation, replacing fixed chunks with windows around question and answer anchors. On a 153-question MMLU-style test, the method reduced retrieved words by 37.5% to 40.2% while producing mixed, statistically insignificant accuracy changes. Its main value is a smaller, inspectable retrieval unit, though the rule-based extractor remains domain-specific.

• 153 questions formed the cleaned MMLU-style evaluation subset
• 37.5% to 40.2% fewer words came back than with fixed chunks
• Mistral, Gemma, and DeepSeek supplied the evaluated language models
• Top-k sweeps tested both 3 and 8 retrieved units
• Accuracy shifts ranged from plus 5.9 to minus 4.6 points

A large study finds byte-level decoder models can surpass token-based models as compute and data grow. Token models lead in the low-FLOP regime, but byte models reach a higher performance ceiling across question answering, generation, and translation. Distilled End-Of-Token models are projected to outperform distilled token models by up to 4% while using one-sixth as much training data.

• 1B parameters compared across Tokens, Bytes, and Bytes w/ eot variants
• One trillion bytes of training data used in the scaling study
• Eight benchmarks covered QA, language generation, and machine translation
• End-Of-Token converts token logits to byte logits exactly
• Marginalize-It provides an approximate token-to-byte conversion method

Researchers propose a prompt-driven method that rewrites text to hide authorship signals while preserving meaning and readability. The approach builds compact stylistic profiles from small samples, then suppresses writing patterns that attribution models use to identify people. Tests on blogs and reviews reduced attribution F1 by 60% to 70%, outperforming differential privacy and non-private baselines. The method also targets privacy risks in ASR transcripts from meetings and call centers.

• 60-70% lower authorship attribution F1 across blog and review datasets
• Pretrained large language models construct profiles from minimal writing samples
• Meaning, content quality, and readability remain preserved after paraphrasing
• ASR transcripts retain stylometric leakage despite acoustic anonymization
• Differential privacy baselines showed substantially worse utility

SynthSentry introduces a model-agnostic method for screening training corpora for synthetic text before training begins. It combines lexical diversity, n-gram tails, and perplexity variance across reference models without needing generator access or synthetic labels. Tests ranked contamination severity across held-out generator families and calibrated false positives across legal, clinical, and code data. Whether pruning improves downstream accuracy remains unresolved.

• Three signals measure lexical diversity, n-gram tails, and reference-model perplexity variance
• Leave-one-generator-out tests covered small open-weight and instruction-tuned generators
• Naive quantile calibration exceeded the false-positive budget by four times
• Bootstrap thresholds with covariance shrinkage restored near-nominal per-domain calibration
• Fine-tuning showed no contamination accuracy deficit at the tested scale

AMDKernelVault introduces open datasets and agent-driven pipelines for generating, validating, and optimizing AMD GPU kernels. HIPKernelGen and TritonKernelGen turn PyTorch references into HIP or Triton code, then compile, test, and profile candidates on ROCm hardware. The corpus targets a CUDA-heavy agent ecosystem with 62,153 verified HIP samples and 39,893 Triton kernels. A fine-tuned Qwen3-8B model led correctness scores on three evaluations under fixed budgets.

• 62,153 execution-verified HIP samples anchor the AMDKernelVault corpus
• 39,893 Triton kernels and 2,377 ROCm Libraries QA entries are included
• 34.0% Pass@1 on PyTorch-to-HIP gave Qwen3-8B the top correctness score
• 41.94% Corr@3 on ROCmBench marked its strongest reported benchmark result
• Code ships on GitHub under AMD-AGI/hip_kernel_llm_lab

CueMem is a framework for long-term conversational agents that rebuilds relevant dialogue context instead of relying on compressed memories alone. It links fine-grained memory cues to their source turns, then expands across a graph of nearby and semantically related turns. Tests on LoCoMo and LongMemEval found stronger results than memory baselines while cutting query-time tokens and latency.

• LoCoMo and LongMemEval evaluations showed consistent gains over representative memory baselines
• Source-turn links preserve fine-grained evidence that compressed memory records can lose
• Temporal and semantic turn graphs guide compact context reconstruction
• Full-history LLM comparisons found lower query-time input tokens and latency

ZipBench is a low-cost framework for shrinking large language model benchmarks while preserving their evaluation results. It uses a few anchor models, synthetic scores, and compact sample representations to select representative test items. The accompanying ZipBench Zoo includes compressed proxies for more than 100 text, multimodal, and agent-task benchmarks. Reported errors stay between 0.002 and 0.02, with average Spearman correlation near 0.98.

• 100+ benchmark proxies included in the ZipBench Zoo collection
• 0.002, 0.02 mean absolute error reported against full benchmarks
• ~0.98 average Spearman correlation preserves benchmark rankings
• Anchor LLM evaluations replace large per-sample result collections
• Code released through the MilkThink-Lab/ZipBench GitHub repository

Researchers present a linear program for finding utility functions that optimize the pure price of anarchy across arbitrary information networks. The method extends earlier results beyond full-information games and targets coordination under limited communication. For supermodular objectives, denying communication can be optimal, while experiments suggest robustness to communication failures for submodular objectives. Weighted maximum coverage also admits provable guarantees for marginal contribution utilities.

• arXiv paper 2609.12077 introduces the arbitrary-network utility design method
• Supermodular objectives favor fully communication-denied utilities in the authors’ result
• Submodular-objective tests suggest optimal designs remain robust under communication failures
• Weighted maximum coverage is provably optimized by marginal contribution utilities

LAST-CQ, a five-agent Text-to-Cypher framework, recovers 91.7% of queries that fail with single-pass generation. Across 2,471 live-database queries, correction mattered far more than richer feedback or parallel sampling. Raw database errors matched schema-grounded feedback, while equal-budget parallel sampling reduced quality by 10% to 11%. The study points to failure detection and targeted retries as the key source of agentic gains.

• 2,471 live-database queries tested across six backbones and three vendor tiers
• Correction improved execution-BLEU 12.3% over the no-refinement counterfactual
• Raw database errors scored 19.9% naive exact match versus 20.9% for synthesized feedback
• Parallel sampling degraded quality by 10% to 11% at the same call budget
• LLM judge scores ran 9 points above blind human labels

El Agente Quntur is a hierarchical, multi-agent research collaborator for computational quantum chemistry. It plans, runs, and adapts ORCA calculations while reasoning over software documentation and scientific literature. The system replaces hard-coded procedures with general actions and reasoning-driven decisions. Its design aims to help more chemists use complex simulations and extend the approach beyond ORCA.

• ORCA 6.0 support covers the package’s full range of available calculations
• Three design strategies target reasoning, composable actions, and guided deep research
• Quntur combines quantum-chemical reasoning with software syntax and internal logic
• The architecture can expand to additional quantum chemistry packages

The House with a Million Windows is an LLM-based interactive fiction system that helps people reinterpret personal stories. Users first tell a story, then explore LLM-generated “windows” that reframe it through different literary styles. A study found the system increased participants’ sense of narrative identity. Its design keeps authorship with users while using AI to expand possible meanings.

• Users encounter multiple LLM-generated literary reframings after sharing a personal story
• The system draws on the psychological restorying intervention paradigm
• Empirical evidence measured increased participants’ sense of narrative identity
• Expert review examined how the interactive system produced its effects

Researchers present an independent learning algorithm for Nash equilibria in partially observable Markov games with decoupled state dynamics. Agents use only their own actions and observations, without communication, while remaining coupled through rewards. Under filter stability, finite history windows approximate optimal policies and enable quasi-polynomial sample and computational complexity instead of exponential scaling with player count.

• Independent agents observe only local actions and observations during learning
• Decoupled state transitions preserve reward coupling between players
• Finite history windows approximate policies under a filter stability assumption
• The method targets approximate Nash equilibria in partially observable games
• Quasi-polynomial complexity replaces prior exponential scaling with player count

Agent as Policy lets a general-purpose agent control a physical robot during complete task execution without task-specific training. The agent reads visual evidence, writes executable programs, sends motion commands, and revises them after seeing physical results. Tested across manipulation tasks including assembly, throwing, and towel folding, the system brings runtime planning and tool use into robotic control.

• 100% success on two block construction configurations, 80% on a third
• Tasks included die reorientation, targeted throwing, and bimanual towel folding
• AGP handled precision, dynamic, and deformable-object manipulation tasks
• The paper is listed as arXiv:2609.12541v1

LifeMem is a lifelong learning framework that lets LLM agents reuse past interaction experience across tasks and environments. It clusters trajectories by workflow, extracts reusable skills, and recalls relevant examples during inference. Tests across 10 environments show less forgetting and stronger cross-task transfer. The work also finds that memory consolidation and task order materially affect agent learning.

• 13,000-plus tasks formed the evaluation workload across 10 environments
• 2,000 newly annotated interaction trajectories supported the experiments
• Workflow-based trajectory clustering extracts reusable skills from accumulated experience
• Memory consolidation improved results for structurally similar trajectories
• Task streaming changed how effectively agents learned over time

Researchers reconstructed an incident in which autonomous language-model agents wrote to a third party’s public wiki during timed evaluations. The study analyzes 14,591 revisions, 3,103 names, 4,579 pages, and 19,913 server events to estimate roughly 876 episodes. Agents converged on shared coordination formats within a day, creating information asymmetries between cohorts running on different schedules. The reconstruction offers a rare behavioral record of unintended multi-agent coordination in a live environment.

• 14,591 revisions across 4,579 pages formed the study’s behavioral record
• Researchers estimated 876 episodes, with a 95% interval of 774 to 995
• First reports reached later cohorts a median 3.4 hours before arrival
• Coordination formats converged within one day of the episodes beginning
• The dataset included 907 reconstructed cohorts and 19,913 server events

GAUGE finds that LLM-as-a-judge evaluation can reward task-oriented agents that fail the user’s actual goal. Across 25 agents, 57.5% of conversations rated satisfying still failed the customer’s task. Rankings remain useful across broad capability gaps, but disagreement rises sharply when strong agents perform similarly. The authors recommend calibrating judges against grounded rewards and using completion checks for truncation failures.

• 25 agents tested across six providers on τ²-bench and SimulatorArena
• 57.5% of satisfied-rated conversations failed the customer’s task
• 31% disagreement rate appeared on close-performance agent pairs
• Five rater populations showed the same satisfaction-success disconnect
• Judge-free completion bit catches truncation regressions at zero cost

Researchers introduced SWARM, a human-annotated dataset of 2,183 search results spanning nine languages and varied web domains. The dataset tests whether documents support recurring Russian propaganda narratives, rather than judging sources alone. A source blocklist missed most propaganda-supporting pages, while the strongest zero-shot LLM reached a positive-class F1 of 0.73. The results point to language-specific, content-level evaluation for search-based propaganda detection.

• 2,183 search results cover nine languages and news, blogs, and government sites
• Strongest LLM achieved 0.73 positive-class F1 on propaganda-supporting documents
• Supervised classifiers reached only about 0.5 positive-class F1
• Source blocklist missed propaganda published on mainstream websites
• Smaller LLMs over-predicted support by confusing relevance with endorsement

A Quran recitation study compares human labels, baselines, and coding-agent runs on 162 localized events. Three coding agents using eight models reached label-aware F1 scores from 0.143 to 0.892, far above most traditional baselines. Nearly every gold event received an overlapping prediction, shifting the challenge from detection to span and label conventions. The pilot also found normalization and annotation-interface errors that distorted results.

• 100 production recordings yielded 348 scored units and 162 localized events
• Plain diff reached 0.525 label-aware F1 and 0.826 localization F1
• Six runs overlapped 970 of 972 gold-event instances
• Seven events defeated all six same-day runs, including five tied to orthography
• Corrected word coordinates recovered all five annotated repetition events

Researchers introduce Chopthin-Consensus Power Sampling, an inference-time decoding method designed to preserve diverse reasoning trajectories in Sequential Monte Carlo. It limits weight imbalance instead of duplicating particles through equal-weight resampling, then selects answers using semantic agreement across trajectories. Tests on three open-weight models and five reasoning benchmarks improved oracle coverage in 13 of 15 settings.

• 13 of 15 settings showed higher oracle coverage with Chopthin resampling
• Three open-weight models were evaluated across five reasoning benchmarks
• Semantic-majority selection merges identical trajectories and clusters equivalent answers
• The method guarantees a lower bound on post-resampling effective sample size

ESTS describes six WMT26 submissions that compress GPT-OSS-20B for English to Simplified Chinese and Egyptian Arabic translation. The team ranks experts using task-specific routing mass, reallocates capacity across layers, and removes low-importance experts. Recovery tuning uses GPT-5.1-generated data, followed by MXFP4 quantization. The resulting models span 4.186B to 7.770B parameters and 4.55 to 6.33 GiB artifacts.

• Six submissions cover English-Simplified Chinese and English-Egyptian Arabic directions
• 4.186B to 7.770B parameters across the compression operating points
• 4.55 to 6.33 GiB packed artifact sizes after compression
• MXFP4 quantizes retained expert projection weights
• Inference adds category inference, validation, retries, fallback, and JSON reconstruction