Daily Summary for 2026-10-01

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: It's the first of October, twenty twenty-six, and this is the day's research.

Jane: 547 new papers came out today.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: We'll take the day in one pass, then pull out the papers we're staying with.

The summary: Tom: Welcome everyone to the research review for the first of October twenty twenty six. Let's dive into today's findings.

Jane: We started with an auditable agentic discovery process for deletion-only tools building on genomic modeling and LLM reasoning.

Lu: Initial steps look at BlockFormer, a transformer method on genomic contact maps to infer structural information.

Meng: That contrasts with understanding computation in multi-task neural networks using the Green's Operator to disentangle pathways.

Lalam: Researchers are also probing how large language models handle reasoning beyond their commitment boundaries via epiphenomenal chain-of-thought processes.

Tom: And they are examining how human label variation relates to formal semantic structure in natural language inference tasks.

Jane: Diffusion language model work shows a dominant self-conditioning direction causes repetition in unconditional continuous diffusion models.

Lu: We are also looking at query-key alignment within LLMs to unlock latent correct answers and zero-shot hallucination detection using lowest span confidence.

Meng: VisionFoundry focused on teaching vision language models visual perception using synthetic images to explore visual data interpretation.

Lalam: They investigated if this method improves the model's visual understanding capabilities. Generalizing the Turing Test to interactive agents was also examined.

Tom: GrepSeek aimed at training search agents for direct corpus interaction, focusing on navigating large text data effectively.

Jane: OctoNest introduced adaptive cross-device execution through stateful control for flexible deployment environments.

Lu: Fork-Think with Confidence explores improving reasoning or confidence levels within a model's decision-making process.

Meng: Removing timing shortcuts in brain-to-text translation refines non-invasive techniques by addressing temporal dependencies in neural data.

Lalam: ETHER introduced aligning emergent communication for hindsight experience replay to better utilize past experiences during learning cycles.

Tom: We are also mitigating memorization in language models by seeking ways to prevent verbatim recall of training data.

Jane: Edit based fingerprints explore unique identifiers by analyzing text edits, suggesting provenance tracking or model drift identification.

Lu: This contrasts with think right which uses adaptive compression to mitigate under-over thinking and improve reasoning accuracy.

Meng: Medrect is a bilingual medical reasoning benchmark designed specifically for error correction within clinical texts.

Lalam: Context aware classification handles sensitive information in online health data, accurately identifying private details in text streams.

Tom: Semantic chunking and the entropy of natural language look at segmenting text based on meaning while measuring its randomness.

Jane: Dataflex presents a unified framework for data-centric dynamic training, adapting models directly from their data sources.

Lu: RA-MoE focuses on routing aligned fine tuning for multilingual adaptation specifically within mixture of experts models.

Meng: This points toward improving cross-lingual performance in complex architectures.

Lalam: Interactor explores agentic reinforcement learning for iterative creation of ad descriptions in sponsored search environments.

Tom: Focusing on interactive generation rather than just static prediction in those environments.

Jane: That concludes our first part of the review for today's research findings. We'll continue next time.

Lu: Thank you all for reviewing these complex topics with me. It was productive today.

Meng: Indeed, the connections between genomic modeling and LLM reasoning are quite fascinating to trace.

Lalam: The focus on agentic behavior across vision and text domains is really expanding the scope of what we can build.

Tom: Precisely, these pieces show how foundational work informs very specialized applications right now.

Jane: We have a lot to unpack in the remaining parts of this review for you all. Stay tuned.

Lu: I look forward to discussing the next set of results with everyone soon.

Meng: Keep an eye out for the next segment; it covers some very deep architectural shifts.

Lalam: It's a rich area, and I think we'll find even more concrete examples there.

Tom: Until then, thank you for tuning in to this research update. Have a great day.

Jane: See you on the next episode for more insights into these developments.

Lu: Goodbye everyone! The research continues tomorrow.

Meng: Until next time for another deep dive into the data.

Lalam: Take care, and keep exploring these fascinating frontiers with me.

Tom: That's all for this first segment of our review today. We'll be back shortly.

Jane: Stay focused on those key findings we discussed, everyone. They are very important for context.

Lu: The interplay between the models and their underlying data structures is what really drives the innovation here.

Meng: It's a complex system, but understanding each component helps map the whole process better.

Lalam: I think we've laid a solid foundation for understanding these cutting-edge areas.

Tom: Absolutely. The work on deletion tools and agentic interaction is where things are moving fast.

Jane: We need to keep pushing the boundaries of what these models can reason about next.

Lu: I agree; the potential for complex agent behavior assessment is huge.

Meng: It requires careful disentanglement, as we saw with the Green's Operator application earlier.

Lalam: Exactly, moving beyond simple classification is the real goal here.

Tom: Let's keep this momentum going into part two of our review.

Jane: I am ready for whatever comes next in the material.

Lu: Ready when you are, Tom. Let's keep exploring the details.

Meng: I am eager to see how these concepts connect in the subsequent sections.

Lalam: Let's keep this conversation flowing with all of you listeners.

Tom: Thank you for your attention throughout this review session today.

Jane: We appreciate your engagement with this advanced research material.

Lu: Keep questioning everything, that is the key to unlocking new insights.

Meng: And remember, every piece of research builds upon the last one we discussed.

Lalam: Indeed, it's a connected tapestry of ideas across many disciplines.

Tom: We'll be back soon with more concrete details on these topics.

Jane: Don't miss the next installment of our research review series.

Lu: Until then, keep your curiosity sharp and your minds open.

Meng: See you all in the next episode for further exploration.

Lalam: Have a wonderful day filled with interesting discoveries!

Tom: Good day to you all listeners. We'll see you soon.

Jane: Keep learning, everyone. That is the most important takeaway from today's review.

Lu: Indeed, the pace of this field is truly accelerating right now.

Meng: The challenges are significant, but the solutions being proposed are very ambitious and exciting.

Lalam: I share that excitement for what these new tools might enable in practice.

Tom: Let's keep analyzing these findings critically as we go on.

Jane: That is the right approach for any serious researcher in this space.

Lu: We will continue to bring you the most accurate and detailed summaries possible.

Meng: Our goal is always to make these complex ideas accessible, step by step.

Lalam: And today we've taken a great first step together on this journey.

Tom: Indeed, thank you for joining us for this comprehensive research review.

Jane: We look forward to continuing this dialogue with you next time.

Lu: Until then, keep pushing the boundaries of what is possible in AI and science.

Meng: Keep an open mind; that is the most valuable tool we possess here.

Lalam: And I hope you find these topics as engaging as we do.

Tom: We’ll be back next time with more research to discuss!

Jane: Until then, stay informed and stay curious about the future of AI.

Lu: Goodbye for now, everyone. Keep up the great work!

Meng: Until our next session, keep exploring these fascinating frontiers.

Lalam: Farewell for today! See you all soon!

Tom: So CombEval showed that high scores on combinatorial tests don't guarantee robust counting abilities, do they?

Jane: Exactly. It means we need to probe deeper into internal logic, not just surface performance.

Lu: That connects to QuantCode, which specialized LLMs for executable trading code using proprietary examples.

Meng: And there's the spike-driven vision-language-action model exploring visual data for action planning in complex environments.

Lalam: I read about how ambiguity in prompts significantly affects LLM performance, linking to zero-label tabular learning.

Tom: Right, and attention mechanism evolution shows trade-offs between efficiency and long-range dependency capture.

Jane: That’s also supported by directed transfer in instruction-tuning mixtures for improved overall performance.

Lu: LatentHarness used counterfactual policy distillation to learn latent actions for memory and reasoning.

Meng: We also saw stress-testing LLM lie detectors revealing that role-play failures cause breakdowns in deception detection.

Lalam: On the Vietnamese side, ViLegalExpert created a benchmark using real consultations for legal Q&A.

Tom: And 4MT-VLM examined how coarse the internal cognitive map is when processing visual and linguistic info.

Jane: There was also work on improving small language models by reusing feedback from larger ones offline and online.

Lu: To fix search algorithms, they looked at making grid beam search less greedy to explore more answers.

Meng: RAIM introduced robust aggregation of inexpensive models specifically for hallucination detection robustness.

Lalam: They also tried compressing looped models using a tilted bowl analogy and defining understanding via bounded Dutch books.

Tom: And Ready2Blend addressed translating natural language instructions into composable alignment prompts.

Jane: It seems like a lot of work connecting these different areas of LLM research today.

Lu: Definitely, it shows the breadth of what we're probing right now.

Meng: Every piece points to needing more sophisticated evaluation and internal structure understanding.

Lalam: True, especially when dealing with ambiguity and complex tasks across modalities.

Tom: So, the takeaway is that surface performance isn't enough; we need deeper structural insights.

Jane: Precisely. The framework for testing those deeper logics is key to our next steps.

Tom: So, the TTLab at Daleel focused on STAR-Ar for argument recognition in Arabic text. It suggests progress in understanding how arguments are structured there.

Jane: That's interesting. And DuplexAct-Bench broadened full-duplex speech evaluation for proactive interaction across different needs.

Lu: I also read about the computational linguistic analysis on Frei.Wild, exploring the link between right-wing rhetoric and rock music using computational methods.

Meng: CATCH was developed as a testbed for reward hacking in reinforcement learning coding environments, probing agent behavior under constraints.

Lalam: There was also work on zero-compute cross-lingual transferability estimation using typological feature proxies to gauge language model translation without high compute.

Tom: MemCodex developed a self-programming hierarchical memory system for language agents to manage and utilize information effectively during operation.

Jane: And synthetic pre-pretraining, while scaling up data improved performance, didn't necessarily translate into gains in grammatical prior compared to smaller models.

Lu: Research on cognitive enhancement suggested rethinking the necessity of role-playing for LLMs when considering their operational needs.

Meng: The investigation into whether computation from earlier problems could assist LLMs in solving novel ones was also conducted.

Lalam: We also saw an arithmetic-dependent rejection bottleneck in the Jev problem, leading to work on verifying rule-governed decisions.

Tom: SEPAL explored separating expert pairs using answer-level fusion to achieve more reliable collaboration between LLMs.

Jane: Today's papers are: From DNA Design to DNA Slimming: Auditable Agentic Discovery of a Deletion-Only Designer.

Lu: BlockFormer uses a transformer to infer biological interactions from genomic contact maps.

Meng: Disentangling Computation in Multi-Task Neural Networks with the Green's Operator separates different computational paths.

Lalam: Beyond the Commitment Boundary probes epiphenomenal chain-of-thought in large reasoning models.

Tom: A Dominant Self-Conditioning Direction Drives Repetition in Unconditional Continuous Diffusion Language Models.

Jane: How Much Human Label Variation Does Formal Semantic Structure Explain? Group-Level Effects and Item-Level Ceilings in NLI.

Lu: Listening to the Wise Few unlocks latent correct answers in large language models using query-key alignment.

Meng: Lowest Span Confidence detects hallucinations by checking the confidence of the lowest span in an LLM's response.

Lalam: VisionFoundry teaches vision-language models visual perception with synthetic images.

Tom: Generalizing the Turing Test to Interactive Agents explores generalizing the Turing test concept for interactive AI agents.

Jane: GrepSeek trains search agents for direct corpus interaction. OctoNest allows adaptive cross-device execution through stateful control.

Lu: Fork-Think with Confidence discusses thinking critically and maintaining confidence when exploring multiple reasoning paths in models.

Meng: Removing Timing Shortcuts Improves Non-Invasive Brain-to-Text by removing timing shortcuts in the process.

Lalam: ETHER aligns emergent communication patterns to improve hindsight experience replay in reinforcement learning.

Tom: Mitigating Memorization In Language Models explores methods to reduce memorization within large language models.

Jane: From Construction to Injection: Edit-Based Fingerprints for Large Language Models uses edit-based fingerprints to manage modifications.

Lu: Think Right learns to mitigate under or overthinking via adaptive, attentive compression.

Meng: MedRECT is a benchmark for bilingual medical reasoning and error correction in clinical texts.

Lalam: Context-Aware Classification and Grading of Sensitive Information in Online Conversational Health Data classifies sensitive health information.

Tom: Semantic Chunking and the Entropy of Natural Language analyzes how semantic chunking affects the randomness of natural language.

Jane: DataFlex provides a unified framework for data-centric dynamic training of large language models.

Lu: RA-MoE fine-tunes mixture-of-experts models for multilingual adaptation by aligning the routing process.

Meng: Interactor uses agentic reinforcement learning to iteratively create ad descriptions in sponsored search.

Lalam: CombEval evaluates combinatorial counting in large language models using a framework designed for that task.

Tom: Using Fine-Tuned LLMs to Identify Indicators of Vulnerability in UK Police Incident Logs finds indicators of vulnerability.

Jane: Frozen Memory Is Not Enough: Rethinking External Memory as Extraction suggests external memory should be treated more like extraction.

Lu: An Empirical Study of Reward Specification and Benchmark Reliability in GRPO-based LLM Unlearning tests unlearning capabilities.

Meng: Argument Structure Prediction in Online Conversations compares different modeling approaches for predicting argument structures.

Lalam: Concept Subspaces Compute Beyond the Logit Lens tests whether concept subspaces can locate representations upstream of the readout layer.

Tom: Working Around the Compute Ceiling: Byte-Exact Memory in Galahad Makes LLM Reading a One-Time Cost.

Jane: Exploring Heterogeneous Model Merging Approach for Complex Knowledge Transfer explores merging different models for complex knowledge transfer.

Lu: QuantCode Model specializes language models specifically for generating executable algorithmic trading code.

Meng: Spike-driven Vision-Language-Action Model is a model driven by spike dynamics.

Lalam: Compact Language, Complex Model Shifts investigates how ambiguity affects language models when the language becomes compact.

Tom: Marginal Response Surface Elicitation for Zero-Label Tabular Learning elicits marginal response surfaces for zero-label tabular learning.

Jane: The Evolution of Attention in Large Language Models reviews the evolution of attention mechanisms and their trade-offs.

Lu: A helps B while B hurts A examines directed transfer when mixing instruction tuning data where A helps B but hurts A.

Meng: LatentHarness learns latent actions for memory and reasoning via counterfactual policy distillation.

Lalam: Stress-Testing LLM Lie Detectors examines role-play failures and spurious correlations in large language models.

Tom: ViLegalExpert is a large benchmark for Vietnamese legal retrieval and question answering from real consultations.

Jane: 4MT-VLM investigates how coarse is the cognitive map in vision-language models.

Lu: Offline Guidance, Online Reasoning reuses LLM feedback for reasoning in smaller language models using offline guidance.

Meng: Making Grid Beam Search Less Greedy aims to improve the exploration strategy of grid beam search.

Lalam: RAIM robustly aggregates inexpensive models to detect hallucinations by aggregating them robustly.

Tom: A Tilted Bowl Is Not a Slippery Slope discusses compressing looped models and tilting the bowl in this context.

Jane: Understanding as No-Arbitrage uses bounded Dutch books as both a definition and training objective for language models.

Lu: Ready2Blend converts natural-language instructions into composable alignment prompts.

Tom: TTLab at Daleel 2026 introduces STAR-Ar for sequence tagging and argument recognition in Arabic.

Jane: DuplexAct-Bench broadens full-duplex speech evaluation toward proactive interaction across diverse behavioral needs.

Lu: Right-Wing Rock or Just Rock? A Computational Linguistic Analysis of Frei.Wild provides a computational linguistic analysis.

Meng: CATCH is a controllable testbed designed to analyze reward hacking in reinforcement learning for coding tasks.

Lalam: Thinking Outside the Box asks whether language models can selectively rely on external guidance when thinking outside the box.

Tom: Zero-Compute Cross-Lingual Transferability Estimation Using Typological Feature Proxies estimates cross-lingual transferability without zero compute.

Jane: Better Supervision Is Nearby uses neighborhood on-policy self-distillation to provide better supervision.

Lu: Drift Inspector explores and measures scientific drift by using atomic contribution claims.

Meng: MemCodex gives language agents self-programming hierarchical memory capabilities for managing information effectively.

Lalam: Synthetic Pre-pretraining Survives Scale, but Not as a Grammatical Prior examines whether synthetic pre-pretraining fails to provide a strong grammatical prior.

Tom: Cognitive Enhancement Rethinking the Necessity of Role-Playing for Large Language Models rethinks the need for role-playing in LLMs.

Jane: Can Computation from Earlier Problems Help LLMs Solve New Ones? investigates whether computation derived from earlier problems can help LLMs solve new ones.

Tom: That concludes our review for today. Next up, we have DNA Design to DNA Slimming: Auditable Agentic Discovery of a Deletion-Only Designer.

Jane: And BlockFormer: Transformer-based inference from genomic contact maps.

Lu: Disentangling Computation in Multi-Task Neural Networks with the Green's Operator.

Meng: Beyond the Commitment Boundary: Probing Epiphenomenal Chain-of-Thought in Large Reasoning Models.

Lalam: A Dominant Self-Conditioning Direction Drives Repetition in Unconditional Continuous Diffusion Language Models.

Tom: How Much Human Label Variation Does Formal Semantic Structure Explain? Group-Level Effects and Item-Level Ceilings in NLI.

Jane: Listening to the Wise Few: Query-Key Alignment Unlocks Latent Correct Answers in Large Language Models.

Lu: Lowest Span Confidence: Zero-Shot Hallucination Detection from a Single LLM Response.

Meng: VisionFoundry: Teaching VLMs Visual Perception with Synthetic Images.

Lalam: Generalizing the Turing Test to Interactive Agents explores generalizing the Turing test concept for interactive AI agents.

Tom: GrepSeek: Training Search Agents for Direct Corpus Interaction.

Jane: OctoNest: Adaptive Cross-Device Execution through Stateful Control.

Lu: Fork-Think with Confidence discusses thinking critically and maintaining confidence when exploring multiple reasoning paths in models.

Meng: Removing Timing Shortcuts Improves Non-Invasive Brain-to-Text.

Lalam: ETHER: Aligning Emergent Communication for Hindsight Experience Replay.

Tom: Mitigating Memorization In Language Models explores methods to reduce memorization within large language models.

Jane: From Construction to Injection: Edit-Based Fingerprints for Large Language Models.

Lu: Think Right: Learning to Mitigate Under-Over Thinking via Adaptive, Attentive Compression.

Meng: MedRECT is a Bilingual Medical Reasoning Benchmark for Error Correction in Clinical Texts.

Lalam: Context-Aware Classification and Grading of Sensitive Information in Online Conversational Health Data.

Tom: Semantic Chunking and the Entropy of Natural Language analyzes how semantic chunking affects language randomness.

Jane: DataFlex: A Unified Framework for Data-Centric Dynamic Training of Large Language Models.

Lu: RA-MoE Routing-Aligned Fine-Tuning for Multilingual Adaptation of Mixture-of-Experts Models.

Meng: Interactor: Agentic RL oriented Iterative Creation for Ad Description Generation in Sponsored Search.

Lalam: CombEval: A Framework for Evaluating Combinatorial Counting in Large Language Models.

Tom: Using Fine-Tuned LLMs to Identify Indicators of Vulnerability in UK Police Incident Logs.

Jane: Frozen Memory Is Not Enough: Rethinking External Memory as Extraction.

Lu: An Empirical Study of Reward Specification and Benchmark Reliability in GRPO-based LLM Unlearning.

Meng: Argument Structure Prediction in Online Conversations: A Comparative Study of Modeling Paradigms and Task Architectures.

Lalam: Concept Subspaces Compute Beyond the Logit Lens: A Weights-Only Test for Locating Representations Upstream of Readout.

Tom: Working Around the Compute Ceiling: Byte-Exact Memory in Galahad Makes LLM Reading a One-Time Cost.

Jane: Exploring Heterogeneous Model Merging Approach for Complex Knowledge Transfer.

Lu: QuantCode Model: Specializing Language Models for Executable Algorithmic Trading Code.

Meng: Spike-driven Vision-Language-Action Model is a vision-language-action model driven by spike dynamics.

Lalam: Compact Language, Complex Model Shifts: How and Where Ambiguity and Underspecification Affect LLMs when the language becomes compact.

Tom: Marginal Response Surface Elicitation for Zero-Label Tabular Learning.

Jane: The Evolution of Attention in Large Language Models reviews the evolution of attention mechanisms and their trade-offs.

Lu: A helps B while B hurts A: directed transfer in instruction-tuning mixture.

Meng: LatentHarness learns latent actions for memory and reasoning via counterfactual policy distillation.

Lalam: Stress-Testing LLM Lie Detectors: Role-Play Failures and Spurious Correlations.

Tom: ViLegalExpert is a Large-Scale Benchmark for Vietnamese Legal Retrieval and Question Answering from Real-World Consultations.

Jane: 4MT-VLM: How Coarse Is a VLMs Cognitive Map?

Lu: Offline Guidance, Online Reasoning: Reusing LLM Feedback for Small Language Models.

Meng: Making Grid Beam Search Less Greedy.

Lalam: RAIM robustly aggregates inexpensive models to detect hallucinations.

Tom: A Tilted Bowl Is Not a Slippery Slope: Compressing Looped Models.

Jane: Understanding as No-Arbitrage: Bounded Dutch Books as a Definition and Training Objective for Language Models.

Lu: Ready2Blend: From Natural-Language Instructions to Composable Alignment Prompts.

Tom: TTLab at Daleel 2026: STAR-Ar, Sequence Tagging for Argument Recognition in Arabic.

Jane: DuplexAct-Bench: Broadening Full-Duplex Speech Evaluation toward Proactive Interaction.

Lu: Right-Wing Rock or Just Rock? A Computational Linguistic Analysis of Frei.Wild.

Meng: CATCH: A Controllable Analysis Testbed for Reward Hacking in Coding RL.

Lalam: Thinking Outside the Box: Can Language Models Rely on External Guidance Selectively?

Tom: Zero-Compute Cross-Lingual Transferability Estimation Using Typological Feature Proxies.

Jane: Better Supervision Is Nearby: Neighborhood On-Policy Self-Distillation.

Lu: Drift Inspector: Exploring and Measuring Scientific Drift with Atomic Contribution Claims.

Meng: MemCodex: Self-Programming Hierarchical Memory for Language Agents.

Lalam: Synthetic Pre-pretraining Survives Scale, but Not as a Grammatical Prior.

Tom: Cognitive Enhancement: Rethinking the Necessity of Role-Playing for Large Language Models.

Jane: Can Computation from Earlier Problems Help LLMs Solve New Ones?

Lu: That concludes our review for today. Next up, we have DNA Design to DNA Slimming: Auditable Agentic Discovery of a Deletion-Only Designer.

Tom: And BlockFormer: Transformer-based inference from genomic contact maps.

Jane: Disentangling Computation in Multi-Task Neural Networks with the Green's Operator.

Lu: Beyond the Commitment Boundary: Probing Epiphenomenal Chain-of-Thought in Large Reasoning Models.

Meng: A Dominant Self-Conditioning Direction Drives Repetition in Unconditional Continuous Diffusion Language Models.

Lalam: How Much Human Label Variation Does Formal Semantic Structure Explain? Group-Level Effects and Item-Level Ceilings in NLI.

Tom: Listening to the Wise Few: Query-Key Alignment Unlocks Latent Correct Answers in Large Language Models.

Jane: Lowest Span Confidence: Zero-Shot Hallucination Detection from a Single LLM Response.

Lu: VisionFoundry: Teaching VLMs Visual Perception with Synthetic Images.

Meng: Generalizing the Turing Test to Interactive Agents explores generalizing the Turing test concept for interactive AI agents.

Lalam: GrepSeek: Training Search Agents for Direct Corpus Interaction.

Tom: OctoNest: Adaptive Cross-Device Execution through Stateful Control.

Jane: Fork-Think with Confidence discusses thinking critically and maintaining confidence when exploring multiple reasoning paths in models.

Lu: Removing Timing Shortcuts Improves Non-Invasive Brain-to-Text.

Meng: ETHER: Aligning Emergent Communication for Hindsight Experience Replay.

Lalam: Mitigating Memorization In Language Models explores methods to reduce memorization within large language models.

Tom: From Construction to Injection: Edit-Based Fingerprints for Large Language Models.

Jane: Think Right: Learning to Mitigate Under-Over Thinking via Adaptive, Attentive Compression.

Lu: MedRECT is a Bilingual Medical Reasoning Benchmark for Error Correction in Clinical Texts.

Meng: Context-Aware Classification and Grading of Sensitive Information in Online Conversational Health Data.

Lalam: Semantic Chunking and the Entropy of Natural Language analyzes how semantic chunking affects language randomness.

Tom: DataFlex: A Unified Framework for Data-Centric Dynamic Training of Large Language Models.

Jane: RA-MoE Routing-Aligned Fine-Tuning for Multilingual Adaptation of Mixture-of-Experts Models.

Lu: Interactor: Agentic RL oriented Iterative Creation for Ad Description Generation in Sponsored Search.

Meng: CombEval: A Framework for Evaluating Combinatorial Counting in Large Language Models.

Lalam: Using Fine-Tuned LLMs to Identify Indicators of Vulnerability in UK Police Incident Logs.

Tom: Frozen Memory Is Not Enough: Rethinking External Memory as Extraction.

Jane: An Empirical Study of Reward Specification and Benchmark Reliability in GRPO-based LLM Unlearning.

Lu: Argument Structure Prediction in Online Conversations: A Comparative Study of Modeling Paradigms and Task Architectures.

Meng: Concept Subspaces Compute Beyond the Logit Lens: A Weights-Only Test for Locating Representations Upstream of Readout.

Lalam: Working Around the Compute Ceiling: Byte-Exact Memory in Galahad Makes LLM Reading a One-Time Cost.

Tom: Exploring Heterogeneous Model Merging Approach for Complex Knowledge Transfer.

Jane: QuantCode Model: Specializing Language Models for Executable Algorithmic Trading Code.

Lu: Spike-driven Vision-Language-Action Model is a vision-language-action model driven by spike dynamics.

Meng: Compact Language, Complex Model Shifts: How and Where Ambiguity and Underspecification Affect LLMs when the language becomes compact.

Lalam: Marginal Response Surface Elicitation for Zero-Label Tabular Learning.

Tom: The Evolution of Attention in Large Language Models reviews the evolution of attention mechanisms and their trade-offs.

Jane: A helps B while B hurts A: directed transfer in instruction-tuning mixture.

Lu: LatentHarness learns latent actions for memory and reasoning via counterfactual policy distillation.

Meng: Stress-Testing LLM Lie Detectors: Role-Play Failures and Spurious Correlations.

Lalam: ViLegalExpert is a Large-Scale Benchmark for Vietnamese Legal Retrieval and Question Answering from Real-World Consultations.

Tom: 4MT-VLM: How Coarse Is a VLMs Cognitive Map?

Jane: Offline Guidance, Online Reasoning: Reusing LLM Feedback for Small Language Models.

Lu: Making Grid Beam Search Less Greedy.

Meng: RAIM robustly aggregates inexpensive models to detect hallucinations.

Lalam: A Tilted Bowl Is Not a Slippery Slope: Compressing Looped Models.

Tom: Understanding as No-Arbitrage: Bounded Dutch Books as a Definition and Training Objective for Language Models.

Jane: Ready2Blend: From Natural-Language Instructions to Composable Alignment Prompts.

Lu: TTLab at Daleel 2026: STAR-Ar, Sequence Tagging for Argument Recognition in Arabic.

Meng: DuplexAct-Bench: Broadening Full-Duplex Speech Evaluation toward Proactive Interaction.

Lalam: Right-Wing Rock or Just Rock? A Computational Linguistic Analysis of Frei.Wild.

Tom: CATCH: A Controllable Analysis Testbed for Reward Hacking in Coding RL.

Jane: Thinking Outside the Box: Can Language Models Rely on External Guidance Selectively?

Lu: Zero-Compute Cross-Lingual Transferability Estimation Using Typological Feature Proxies.

Meng: Better Supervision Is Nearby: Neighborhood On-Policy Self-Distillation.

Lalam: Drift Inspector: Exploring and Measuring Scientific Drift with Atomic Contribution Claims.

Tom: MemCodex: Self-Programming Hierarchical Memory for Language Agents.

Jane: Synthetic Pre-pretraining Survives Scale, but Not as a Grammatical Prior.

Lu: Cognitive Enhancement: Rethinking the Necessity of Role-Playing for Large Language Models.

Meng: Can Computation from Earlier Problems Help LLMs Solve New Ones?

Lalam: That concludes our review for today. Tune in next time for DNA Design to DNA Slimming: Auditable Agentic Discovery of a Deletion-Only Designer.

Tom: And BlockFormer: Transformer-based inference from genomic contact maps.

Jane: Disentangling Computation in Multi-Task Neural Networks with the Green's Operator.

Lu: Beyond the Commitment Boundary: Probing Epiphenomenal Chain-of-Thought in Large Reasoning Models.

Meng: A Dominant Self-Conditioning Direction Drives Repetition in Unconditional Continuous Diffusion Language Models.

Lalam: How Much Human Label Variation Does Formal Semantic Structure Explain? Group-Level Effects and Item-Level Ceilings in NLI.

Tom: Listening to the Wise Few: Query-Key Alignment Unlocks Latent Correct Answers in Large Language Models.

Jane: Lowest Span Confidence: Zero-Shot Hallucination Detection from a Single LLM Response.

Lu: VisionFoundry: Teaching VLMs Visual Perception with Synthetic Images.

Meng: Generalizing the Turing Test to Interactive Agents explores generalizing the Turing test concept for interactive AI agents.

Lalam: GrepSeek: Training Search Agents for Direct Corpus Interaction.

Tom: OctoNest: Adaptive Cross-Device Execution through Stateful Control.

Jane: Fork-Think with Confidence discusses thinking critically and maintaining confidence when exploring multiple reasoning paths in models.

Lu: Removing Timing Shortcuts Improves Non-Invasive Brain-to-Text.

Meng: ETHER: Aligning Emergent Communication for Hindsight Experience Replay.

Lalam: Mitigating Memorization In Language Models explores methods to reduce memorization within large language models.

Tom: From Construction to Injection: Edit-Based Fingerprints for Large Language Models.

Jane: Think Right: Learning to Mitigate Under-Over Thinking via Adaptive, Attentive Compression.

Lu: MedRECT is a Bilingual Medical Reasoning Benchmark for Error Correction in Clinical Texts.

Meng: Context-Aware Classification and Grading of Sensitive Information in Online Conversational Health Data.

Lalam: Semantic Chunking and the Entropy of Natural Language analyzes how semantic chunking affects language randomness.

Tom: DataFlex: A Unified Framework for Data-Centric Dynamic Training of Large Language Models.

Jane: RA-MoE Routing-Aligned Fine-Tuning for Multilingual Adaptation of Mixture-of-Experts Models.

Lu: Interactor: Agentic RL oriented Iterative Creation for Ad Description Generation in Sponsored Search.

Meng: CombEval: A Framework for Evaluating Combinatorial Counting in Large Language Models.

Lalam: Using Fine-Tuned LLMs to Identify Indicators of Vulnerability in UK Police Incident Logs.

Tom: Frozen Memory Is Not Enough: Rethinking External Memory as Extraction.

Jane: An Empirical Study of Reward Specification and Benchmark Reliability in GRPO-based LLM Unlearning.

Lu: Argument Structure Prediction in Online Conversations: A Comparative Study of Modeling Paradigms and Task Architectures.

Meng: Concept Subspaces Compute Beyond the Logit Lens: A Weights-Only Test for Locating Representations Upstream of Readout.

Lalam: Working Around the Compute Ceiling: Byte-Exact Memory in Galahad Makes LLM Reading a One-Time Cost.

Tom: Exploring Heterogeneous Model Merging Approach for Complex Knowledge Transfer.

Jane: QuantCode Model: Specializing Language Models for Executable Algorithmic Trading Code.

Lu: Spike-driven Vision-Language-Action Model is a vision-language-action model driven by spike dynamics.

Meng: Compact Language, Complex Model Shifts: How and Where Ambiguity and Underspecification Affect LLMs when the language becomes compact.

Lalam: Marginal Response Surface Elicitation for Zero-Label Tabular Learning.

Tom: The Evolution of Attention in Large Language Models reviews the evolution of attention mechanisms and their trade-offs.

Jane: A helps B while B hurts A: directed transfer in instruction-tuning mixture.

Lu: LatentHarness learns latent actions for memory and reasoning via counterfactual policy distillation.

Meng: Stress-Testing LLM Lie Detectors: Role-Play Failures and Spurious Correlations.

Lalam: ViLegalExpert is a Large-Scale Benchmark for Vietnamese Legal Retrieval and Question Answering from Real-World Consultations.

Tom: 4MT-VLM: How Coarse Is a VLMs Cognitive Map?

Jane: Offline Guidance, Online Reasoning: Reusing LLM Feedback for Small Language Models.

Lu: Making Grid Beam Search Less Greedy.

Meng: RAIM robustly aggregates inexpensive models to detect hallucinations.

Lalam: A Tilted Bowl Is Not a Slippery Slope: Compressing Looped Models.

Tom: Understanding as No-Arbitrage: Bounded Dutch Books as a Definition and Training Objective for Language Models.

Jane: Ready2Blend: From Natural-Language Instructions to Composable Alignment Prompts.

Lu: TTLab at Daleel 2026: STAR-Ar, Sequence Tagging for Argument Recognition in Arabic.

Meng: DuplexAct-Bench: Broadening Full-Duplex Speech Evaluation toward Proactive Interaction.

Lalam: Right-Wing Rock or Just Rock? A Computational Linguistic Analysis of Frei.Wild.

Tom: CATCH: A Controllable Analysis Testbed for Reward Hacking in Coding RL.

Jane: Thinking Outside the Box: Can Language Models Rely on External Guidance Selectively?

Lu: Zero-Compute Cross-Lingual Transferability Estimation Using Typological Feature Proxies.

Meng: Better Supervision Is Nearby: Neighborhood On-Policy Self-Distillation.

Lalam: Drift Inspector: Exploring and Measuring Scientific Drift with Atomic Contribution Claims.

Tom: MemCodex: Self-Programming Hierarchical Memory for Language Agents.

Jane: Synthetic Pre-pretraining Survives Scale, but Not as a Grammatical Prior.

Lu: Cognitive Enhancement: Rethinking the Necessity of Role-Playing for Large Language Models.

Meng: Can Computation from Earlier Problems Help LLMs Solve New Ones?

Lalam: That concludes our review for today. Tune in next time for DNA Design to DNA Slimming: Auditable Agentic Discovery of a Deletion-Only Designer.