Daily Summary for 2026-10-18

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

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

Jane: 299 new papers came out today.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

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

Jane: Start us off with Inference-Time Steering Across Scaffolds and Epitopes for Sample-Efficient De Novo Nanobody Design.

The summary: Tom: Welcome. This is the eighteenth of October, twenty twenty six.

Jane: We started with Harish Haresamudram Wanting Mao Aditya Radhakrishnan Maxwell A Xu and James M Rehg on VarMAE.

Lu: VarMAE uses a variable input masked autoencoder to learn one representation across different sampling rates and temporal contexts.

Meng: It achieves this by creating two views of the same signal using a cross-resolution masking objective called AKMPC.

Lalam: This shows a single foundation model can achieve broad transfer across sensing configurations even with fixed pretraining constraints.

Tom: Then there is VQIL developing a trainable quantum walk for seed-conditioned ranking on graphs.

Jane: VQIL uses quantum interference to rank nodes based on their response strength to a perturbation.

Lu: This shows significant improvements over classical models in both power-flow and protein-protein interaction network scenarios.

Meng: Taeyoung Yu Seonbin Jo Jiwon Kim Junyoung Byun created SeoulMMOD for origin-destination flow forecasting.

Lalam: SeoulMMOD addresses benchmark limitations with multimodal, multi-year data for citywide mobility predictions.

Tom: Mason Minot Simon Friedensohn Cédric R Weber proposed dynamic budget allocators for generative condition selection.

Jane: This is for de novo nanobody design and empirically showed nearly five times the yield per GPU-hour.

Lu: Sparse-TD-JEPA introduces an l1 sparsity penalty to improve zero-shot reinforcement learning performance.

Meng: It concentrates reward structure into a smaller latent subspace, boosting performance in that area.

Lalam: Arthur da Cunha Kasper Green Larsen Liang-Yu Zou's AKMPC study quantifies how boosting weak hypotheses depends on the base class structure.

Tom: They used a gamma-VC dimension to sharpen the link between this and the classic VC dimension for learning.

Jane: This gives us a more precise tool to understand when boosting will actually work well beyond general bounds.

Lu: Richard Cornelius Suwandi Feng Yin Kevin Murphy's Kernaut tackles designing good kernels by treating design as open-ended discovery.

Meng: Agents discover reusable inductive biases in these kernels, and a human-refined version improved predictive error by five point seven percent.

Lalam: Van Dai Do Huu Hiep Nguyen Minh Hoang Nguyen Manh Nguyen Hung Le Deakin's LOTUS introduces forecast grounding for time-series language models.

Tom: This requires the model to predict future evolution before adapting it to downstream tasks with few trainable parameters.

Jane: Rishabh Tiwari Kusha Sareen Lakshya A Agrawal's FST framework separates parameters into slow and fast weights.

Lu: Slow weights retain general knowledge while fast context weights learn task-specific information, reducing catastrophic forgetting.

Meng: Gated Adaptive Positional Encoding addresses instability in large language models handling long sequences.

Lalam: This introduces a content-aware bias into the attention mechanism to suppress irrelevant distant context and protect salient tokens.

Tom: The GAPE framework uses query dependent and key dependent gates for context control.

Jane: That contracts unnecessary context while keeping distant tokens accessible in standard attention.

Lu: Preping builds procedural memory via a proposer guided framework using synthetic tasks.

Meng: The proposer generates tasks conditioned on a structured control state before execution.

Lalam: This reduces deployment costs compared to online construction methods significantly.

Tom: Learning to Pose Problems reasons about problem directions before synthesizing data.

Jane: It adapts difficulty based on the solver's ability and uses consistency rewards to guide it.

Lu: EconCausal tackles cause and effect by using a benchmark from over 2595 economic studies.

Meng: The core issue is inferring effect direction given context change, which models struggle with.

Lalam: Top models hit 84% accuracy in single contexts but drop to 74.75% overall.

Tom: Accuracy plummets to just 44.27% when source and target input signs disagree.

Jane: Accuracy falls below sixty percent when source signs are corrupted, showing input sensitivity.

Lu: CoTS reframes reasoning as continuous allocation between commitment and soft exploration at each step.

Meng: It uses competition between candidates to dynamically balance these modes across benchmarks.

Lalam: PISO combines partial gradients with a running residual estimate for sketch iterations.

Tom: This provides convergence guarantees based on residual variation instead of uniform bounds.

Jane: AdaSwitch adapts by switching between offline and online oracle methods for history dependent problems.

Lu: It offers explicit performance guarantees that tighten as prediction errors decrease or optimum increases.

Meng: UniFGVC avoids overfitting in fine-tuning by using a training free retrieval paradigm.

Lalam: It generates structured text descriptions to pinpoint visual attributes for classification tasks.

Tom: OrBIT learns reusable local geometry from orbit dynamics to guide embedding compression effectively.

Jane: This drastically reduces storage footprint while maintaining high performance metrics.

Lu: OrBIT reaches thirty seven point nine times compression on GPT-2 and twenty three times on 7B models.

Meng: The global reconstruction residual decides where the fixed coding budget is spent elegantly.

Lalam: Redundant overlapping charts allow local errors to compensate after gluing together this structure guided compression.

Tom: Another area is temporal fidelity in video models like Moment-Video.

Jane: That benchmark shows errors persist on longer videos even when sampling covers all events.

Lu: On the agent side, Life-Harness adapts the interface instead of updating model parameters for LLM agents.

Meng: This adaptation can lead to accuracy gains exceeding Meta-Harness significantly for models like Qwen3-4B-Instruct.

Lalam: MOCHA optimizes agent skills using Chebyshev scalarization to improve both accuracy and economy simultaneously.

Tom: Agents can escape their initial seed skill by discovering trade-offs that average agents ignore.

Jane: BankerToolBench creates a benchmark for end-to-end analytical workflows in investment banking tasks.

Lu: Testing frontier models showed GPT-5.4 failed nearly half the rubric criteria for client readiness.

Meng: This points to major problems like breakdowns in cross-artifact consistency during complex tasks.

Lalam: EVALCAST helps study how much can be inferred from natural language descriptions of tasks before experiments.

Tom: AEGIS operationalizes regulatory concepts by monitoring models under standards like the FDA and EU AI Act.

Jane: AEGIS correctly identified deployable decisions while flagging critical states where no model existed.

Lu: Certified Robustness via Sparse Second-Order Cone Relaxations provides stronger certificates than simpler linear methods.

Meng: Masked Attention Pooling is invariant to padding and permutation under ideal real arithmetic conditions.

Lalam: This contrasts with BankerToolBench where consistency breaks down during agent execution.

Tom: Variable Input Masked Autoencoders learns a single representation across different sampling rates for sensor data.

Jane: VQIL uses quantum walks to rank nodes, identifying which components drive or absorb a system's response.

Lu: SeoulMMOD offers comprehensive data for citywide multimodal mobility in Seoul forecasting across multiple modes.

Meng: Inference-Time Steering Across Scaffolds and Epitopes for Sample-Efficient De Novo Nanobody Design reconfigures condition selection.

Lalam: Evaluating Sequence Assembly Strategies for Differentially Private Synthetic Time-Series Forecasting shows utility depends on assembly by the forecaster.

Tom: Lotus introduces forecast grounding, a pretraining principle requiring time-series representations to preserve future evolution information.

Jane: CSWinUNETR proposes a general backbone for segmenting thin anatomical structures in medical images using cross-shaped stripe self-attention.

Lu: Learning, Fast and Slow: Towards LLMs That Adapt Continually separates task-specific information into fast and slow weights.

Meng: GAPE is a drop-in augmentation for positional encodings that uses query-dependent gates to suppress irrelevant long-range context.

Lalam: PREPING builds agent memory through self-generated synthetic practice guided by a proposer state.

Tom: The Minimax Rate of Perturbed Second-Order Calibration characterizes the rate of estimating second-order calibration error.

Jane: Towards Better Aligned LLM Judges for Radiology Report Evaluation across Modalities and Anatomies evaluates different metrics to find robust judges.

Lu: The Reasoning Trap argues that improving logical reasoning enables deeper situational awareness through self-examination.

Meng: A Context-Aware Economic Reasoning Benchmark for Large Language Models tests causal effects under different institutional contexts.

Lalam: AdaSwitch adaptively switches between offline and online oracles based on prediction error to optimize performance.

Tom: UniFGVC frames few-shot visual classification as a multimodal retrieval problem using attribute-aware captions.

Jane: Partially Informed Stochastic Optimization via Zeroth-Order Residual Tracking improves optimization with partial gradient access.

Lu: Decentralized Bandit Learning of Markov Perfect Equilibria proves fully decentralized agents can converge to a stationary equilibrium.

Meng: Efficient Learning of Survival Analysis Models Under Outcome-Dependent Truncation and Censoring uses first-order optimization.

Lalam: CoTS proposes a training-free framework that dynamically balances discrete commitment and soft exploration at the token level during reasoning.

Tom: Learning Traffic Flow Dynamics with Stochastic Physics-Informed Neural Cellular Automata learns traffic flow while preserving physical constraints.

Jane: TR-PTQ introduces an integer-only formulation to eliminate the need for floating point hardware units in quantization.

Lu: Executing Causal Structure Learning with Linear-Attention Transformers shows a fixed transformer block can execute continuous causal discovery updates.

Meng: Neuromotor Hierarchy Network Physiological Inductive Biases for Robust Generalization in sEMG Decoding learns a compact latent neuromotor state.

Lalam: PRISM recovers per-query signal from a corpus by learning with privileged information from sibling graphs.

Tom: KACE decides how much inference to spend and which learned knowledge to show based on problem difficulty tiers.

Jane: C-MIG rewards intermediate steps in clinical diagnosis by measuring information gain from a frozen reference model.

Lu: Ideological Bias in LLMs' Economic Causal Reasoning systematically evaluates predictions under pro-government versus market perspectives.

Meng: Reward Design for Physical Reasoning in Vision-Language Models compares external and internal attention rewards to determine the best signal.

Lalam: Local Timescale Gates for Timescale-Robust Continual Spiking Neural Networks introduces a neuron model combining dual time constants with adaptive gating.

Tom: Uncovering Grounding IDs shows external visual cues help LVLMs form more accurate connections between visual and textual information.

Jane: Towards Reliable Long-Context Reasoning: Detecting Unanswerable Questions via FactGuard generates evidence-grounded questions to improve reliability.

Lu: Scaling Multimodal Search and Recommendation with Small Language Models demonstrates small models can support multimodal search effectively.

Meng: Forget About It investigates whether simple in-training parameter interpolation can mitigate catastrophic forgetting without complex replay mechanisms.

Lalam: LLM Evaluation is a Forecasting Problem suggests model performance should be predicted by time-series forecasters applied to historic performance.

Tom: PGSB identifies shared directions among task-specific LoRA updates to improve model merging efficiency.

Jane: Orbit Structure-Guided Embedding Compression learns reusable local geometry from orbit dynamics to constrain shared codewords for massive compression.

Lu: Distilling Graph Geometry: Knowledge Gap from GNNs to MLPs proposes a distillation framework guided by Ollivier-Ricci curvature.

Meng: KNOU-Flood advances global streamflow forecasting with lead-aware probabilistic machine learning extending open recipes for flood forecasting.

Lalam: Activation Steering studies whether injecting task-specific steering vectors into LLM activations can induce broadly unsafe behavior across domains.

Tom: Momentary Visual Events Diagnosing Temporal Fidelity of Video MLLMs on Momentary Visual Events introduces Moment-Video.

Jane: Adapting the Interface, Not the Model: Runtime Harness Adaptation for Deterministic LLM Agents proposes Life-Harness.

Lu: MOCHA Multi-Objective Chebyshev Annealing for Agent Skill Optimization uses Chebyshev scalarization to select skills improving the Pareto front.

Meng: BankerToolBench Evaluating AI Agents in End-to-End Investment Banking Workflows benchmarks agents on complex, multi-file analytical workflows.

Lalam: AEGIS An Operational Infrastructure for Post-Market Governance of Adaptive Medical AI presents a governance framework for safely managing continuous learning.

Tom: Certified Robustness via Sparse Second-Order Cone Relaxations provides stronger certificates than simpler linear methods.

Jane: Masked Attention Pooling is invariant to padding and permutation under ideal real arithmetic conditions.

Lu: Variable Input Masked Autoencoders learns a single representation across different sampling rates for sensor data.

Meng: VQIL uses quantum walks to rank nodes, identifying which components drive or absorb a system's response.

Lalam: SeoulMMOD offers comprehensive data for citywide multimodal mobility in Seoul forecasting across multiple modes.

Tom: Inference-Time Steering Across Scaffolds and Epitopes for Sample-Efficient De Novo Nanobody Design reconfigures condition selection.

Jane: Evaluating Sequence Assembly Strategies for Differentially Private Synthetic Time-Series Forecasting shows utility depends on assembly by the forecaster.

Lu: Lotus introduces forecast grounding, a pretraining principle requiring time-series representations to preserve future evolution information.

Meng: CSWinUNETR proposes a general backbone for segmenting thin anatomical structures in medical images using cross-shaped stripe self-attention.

Lalam: Learning, Fast and Slow: Towards LLMs That Adapt Continually separates task-specific information into fast and slow weights.

Tom: Remember to Forget Gated Adaptive Positional Encoding proposes GAPE, a drop-in augmentation for positional encodings that uses query-dependent gates.

Jane: PREPING builds agent memory through self-generated synthetic practice guided by a proposer state.

Lu: The Minimax Rate of Perturbed Second-Order Calibration characterizes the rate of estimating second-order calibration error.

Meng: Towards Better Aligned LLM Judges for Radiology Report Evaluation across Modalities and Anatomies evaluates different metrics to find robust judges.

Lalam: The Reasoning Trap argues that improving logical reasoning enables deeper situational awareness through self-examination.

Tom: EconCausal introduces a benchmark to test if LLMs can infer causal effects under different institutional and market contexts.

Jane: AdaSwitch adaptively switches between offline and online oracles based on prediction error to optimize performance.

Lu: UniFGVC frames few-shot visual classification as a multimodal retrieval problem using attribute-aware captions.

Meng: Partially Informed Stochastic Optimization via Zeroth-Order Residual Tracking improves optimization with partial gradient access.

Lalam: Decentralized Bandit Learning of Markov Perfect Equilibria proves fully decentralized agents can converge to a stationary equilibrium.

Tom: Efficient Learning of Survival Analysis Models Under Outcome-Dependent Truncation and Censoring uses first-order optimization.

Jane: CoTS proposes a training-free framework that dynamically balances discrete commitment and soft exploration at the token level during reasoning.

Lu: Learning Traffic Flow Dynamics with Stochastic Physics-Informed Neural Cellular Automata learns traffic flow while preserving physical constraints.

Meng: TR-PTQ introduces an integer-only formulation to eliminate the need for floating point hardware units in quantization.

Lalam: Executing Causal Structure Learning with Linear-Attention Transformers shows a fixed transformer block can execute continuous causal discovery updates.

Tom: Neuromotor Hierarchy Network Physiological Inductive Biases for Robust Generalization in sEMG Decoding learns a compact latent neuromotor state.

Jane: PRISM recovers per-query signal from a corpus by learning with privileged information from sibling graphs.

Lu: KACE decides how much inference to spend and which learned knowledge to show based on problem difficulty tiers.

Meng: C-MIG rewards intermediate steps in clinical diagnosis by measuring information gain from a frozen reference model.

Lalam: Ideological Bias in LLMs' Economic Causal Reasoning systematically evaluates predictions under pro-government versus market perspectives.

Tom: Reward Design for Physical Reasoning in Vision-Language Models compares external and internal attention rewards to determine the best signal.

Jane: Local Timescale Gates for Timescale-Robust Continual Spiking Neural Networks introduces a neuron model combining dual time constants with adaptive gating.

Lu: Uncovering Grounding IDs shows external visual cues help LVLMs form more accurate connections between visual and textual information.

Lucky paper: 2610.13844: Tom: So, we’re diving into Inference-Time Steering Across Scaffolds and Epitopes for Sample-Efficient De Novo Nanobody Design today.

Jane: This paper by Mason Minot et al. is focused on how to make de novo nanobody design much more sample efficient by treating condition selection as a resource allocation problem.

Lu: They propose two dynamic allocators, one joint Global Sequential Monte Carlo formulation and one decoupled approach using Batched Successive Elimination.

Meng: The core idea is that upfront commitment to choices like scaffold identity wastes compute because the yield changes across those conditions.

Tom: So they recast condition selection as an inferencetime resource-allocation problem instead of a fixed choice before training begins.

Jane: They use a joint SMC formulation where the budget allocation happens during global systematic resampling when unpromising particles are extinguished and successful ones replicate across condition boundaries.

Lu: And for the decoupled approach, they use Batched Successive Elimination with a modification where budgets are distributed evenly among surviving arms.

Meng: That modification involves ranking arms by their 90th percentile of their terminal validation scores and enforcing a deterministic fifty percent elimination rate at the end of each round.

Tom: They apply inference-time steering methods like Feynman-Kac steering to sample from these reward-tilted distributions by actively cloning promising trajectories.

Jane: This involves interleaving standard denoising steps with guidance events that evaluate particle quality and discard low scoring states.

Lu: The paper shows that while steering collapses paired CDR-H3 Calpha RMSD within individual trajectories, this loss of variation is localized to single runs.

Meng: But across independent seeds, steering recovers structural variation by exploring distinct, high-reward conformational basins.

Tom: That suggests steering isn't inherently sacrificing global diversity; it points to the need to distribute compute across multiple starting conditions.

Jane: They also found that the joint SMC allocation maximized raw yield in the PD-L1 binder design experiment.

Lu: However, they noted a vulnerability in joint SMC allocation because it allocated fifty-three percent of its particles to the 8z8v-derived framework due to high ESMFold2 ipTM.

Meng: But this yielded zero terminal AF2-M passes because of pose divergence, which is a major problem.

Tom: That’s a huge red flag about reward hacking when you rely too heavily on one metric like ipTM or ipTM scores.

Jane: The paper validates candidates using three metrics: inter-chain predicted alignment error ninety and the self-consistent root-mean-square deviation of the binder under one point five angstrom.

Lu: They concluded that dynamic resource allocation paired with inference-time steering converts static constraints into active optimization variables.

Tom: That means they increase in silico nanobody yield without sacrificing structural diversity, which is what we need.

Jane: The authors also pointed out that high ipTM scores do not guarantee in vitro affinity or stability, meaning the model artifacts are still a risk.

Lu: Adaptation of these allocators to wet-lab-correlated oracles is the primary direction they suggest for future work.

Meng: That means they need to connect this computational steering back to real lab results to really prove its value in practice.

Tom: It sounds like a very clever way to manage the massive compute cost inherent in de novo design.

Lucky paper: 2610.10222: Tom: Moving on to our next paper, we've got Evaluating Sequence Assembly Strategies for Differentially Private Synthetic Time-Series Forecasting.

Jane: This work is really about treating the sequence assembly process as an explicit design stage in forecasting pipelines using differentially private synthetic time series.

Lu: The researchers systematically varied overlap rates and window weighting schemes while keeping the generator fixed to evaluate the resulting sequences based on boundary continuity, statistical fidelity, and Train-on-Synthetic-Test-on-Real utility.

Meng: They measure boundary continuity with a diagnostic called Jjump, which targets discontinuities at assembly junctions.

Lalam: Proposition one shows that gradual weight changes limit the contribution of cross window mismatch while variation within a window remains part of the assembled signal.

Tom: Temporal fidelity is assessed by computing Delta ACF24h equals ACF24h of X-hat minus ACF24h of X.

Jane: And downstream utility is measured using the TSTR protocol where a forecaster trains on the assembled synthetic data and is evaluated on held out real observations.

Lu: They found forecaster dependent assembly regularities, showing recurrent models favor intermediate to high overlap, while ARIMA and Prophet favor zero or low overlap.

Meng: LightGBM showed stronger dataset dependence across all assembly settings in their complete five forecaster grids.

Lalam: Matched Train-on-Real-Test-on-Real TRTR references contextualize utility relative to real training, showing the lowest MAE configurations identify candidate settings for the target forecasting task.

Tom: This means sequence assembly is a forecaster aware model selection problem, giving practical guidance for prioritizing candidate configurations before target task validation.

Jane: The operational validation in electricity load forecasting showed that these same broad forecaster dependent assembly tendencies persisted in practice.

Lu: LSTM and RNN achieve their lowest MAE at seventy-five percent and eighty-five point seven percent overlap respectively, consistent with the higher overlap preference of recurrent forecasters.

Meng: ARIMA and Prophet attain their minima at low or zero overlap, which aligns with their public dataset tendencies, but LightGBM still shows task specific behavior.

Lalam: Downstream TSTR utility provides the basis for configuration selection while continuity and fidelity diagnostics offer complementary interpretation and matched TRTR provides the corresponding real training reference.

Lucky paper: 2610.18877: Tom: Alright, let's turn to Lotus, which is paper number three of five this episode.

Jane: It’s about forecast grounding and how language-aligned time-series representations need to predict future evolution before they get adapted for downstream tasks.

Lu: The core idea is that this pretraining constraint encourages the representation supplied to the LLM to preserve future relevant temporal information instead of just being an extra downstream task.

Tom: They implement this by using a temporal spectral patch representation, combining raw time-series patches with their real Fourier coordinates before mapping them into the LLM embedding space.

Jane: That means for a patch xc,n, they use the raw patch as the temporal view and its real FFT for the spectral view.

Lu: These views are embedded separately by learnable projections Wt and Wf, then fused by Wfuse to create a shared representation hc,n.

Tom: The LOTUS architecture divides each time-series channel into non-overlapping patches that map into soft tokens in the LLM embedding space.

Jane: They have three phases: Phase one is the Forecast Grounding Pretraining, Phase two is Downstream Adaptation, and then they adapt the encoder and LLM using language supervision alone.

Lu: During pretraining, language supervision trains textual outputs while next-patch prediction constrains the LLM's hidden states to retain future relevant information.

Tom: After pretraining, they remove the next-patch decoder and adapt only with language supervision for the downstream task adaptation.

Jane: On Qwen3-0 point 6B, LOTUS achieves an overall accuracy of fifty-nine point nine one percent, which is four point five eight percent above TS-Reasoner-7B-FT while being fourteen times smaller in backbone size.

Lu: The representation analysis confirms that LOTUS uses the numerical input and preserves temporal structure in its learned patch representations very well.

Tom: Ablation studies confirm the benefits, showing that removing the frequency branch lowers six task average accuracy by three point eight percent in Phase two alone and four point seven percent in both phases.

Jane: Input diagnostics show that accuracy on HAR falls from fifty-eight point four percent to between twenty-six point nine and thirty-two point two percent when the numerical input is corrupted with noise or zeroed out, which demonstrates the predictions depend materially on the numerical input itself.

Lu: The paper establishes forecast grounding as a promising pretraining principle enabling compact time-series language models to outperform models twenty-five to fifty times their size on TSRBench.

Tom: They show forecasting as a pretraining objective works even when downstream tasks only require language outputs.

Lucky paper: 2606.19824: Tom: So we're looking at CSWinUNETR, which focuses on segmentation of thin, tortuous anatomical structures in medical images.

Jane: It uses a cross-shaped stripe self-attention along with detail enhanced multi-scale self-attention and sparse control dynamic snake convolution.

Lu: The architecture has three core components: Shifted CSWin Self-Attention for orientation awareness, Detail Enhanced MS-MHSA for fine details, and SDSConv to encourage continuity along tortuous trajectories.

Meng: The Shifted CSWin Self-Attention partitions attention heads into vertical and horizontal stripes and applies a cyclic shift in the second block to alleviate stripe isolation artifacts.

Lalam: The Detail Enhanced MS-MHSA module constructs several lightweight depthwise convolution branches, including an axial strip branch and a semantic context branch with average pooling followed by a three times three DWC.

Tom: And then SDSConv addresses anisotropic evidence aggregation by deforming an axial convolutional kernel using location adaptive offsets instead of regressing independent offsets at each spatial location.

Jane: That dense snake like kernel is constructed by interpolating those offsets along the kernel using Gaussian radial basis functions or RBFs.

Lu: Experimentally, CSWinUNETR consistently outperforms state of the art methods across four heterogeneous benchmarks in ophthalmology, neurovascular imaging, and dermatology.

Meng: Table three shows that adding the cyclic shift, MS-MHSA, and SDSConv leads to consistent stepwise improvements across all evaluation metrics.

Lalam: The qualitative results show CSWinUNETR more faithfully reconstructs fine branches and terminal structures while reducing false positives in non target regions.

Tom: It seems like the architectural inductive biases are more responsible for the gains than just tuning hyperparameters on a specific dataset.

Jane: It is interesting how it handles continuity even when the structure looks locally disrupted due to limited spatial resolution or motion artifacts.

Lu: The paper emphasizes that accurate segmentation requires both orientation aware long range information propagation and anisotropic trajectory aligned integration of local evidence.

Tom: We also saw comparisons with baselines in Table one and Table two showing CSWinUNETR consistently wins across all four benchmarks.

Jane: That unified setting indicates the performance gains are more plausibly attributable to CSWinUNETR’s architectural inductive biases than to extensive dataset specific hyperparameter tuning.

Lu: The paper fixed the base embedding dimension at C0 equals forty-eight and set the number of attention heads depending on whether it's a 2D or three dee input.

Meng: For SDSConv, the kernel length P was set to (eleven nine seven and five) across different stages from shallow to deep.

Lalam: The evaluation metrics used include Dice and clDice overlap metrics as well as topology error Betti and boundary distance error HD95.

Lucky paper: 2605.12484: Tom: We're moving into our fifth paper of five today, and we're looking at Learning, Fast and Slow: Towards LLMs That Adapt Continually by Rishabh Tiwari, Kusha Sareen, and Lakshya A Agrawal.

Jane: This paper introduces the Fast-Slow Training framework where model parameters are treated as slow weights and context is treated as fast weights.

Lu: The slow parametric component, denoted by θ, represents the model weights which are expensive to update and persist across tasks.

Meng: The fast textual component, denoted by ϕ, represents prompts and context which can be changed cheaply to capture task-level adaptation without permanently modifying the model.

Lalam: This division of labor allows the fast weights to learn from textual feedback while allowing slow weights to stay closer to the base model and persist general reasoning behaviors.

Tom: So, how does this framework actually work in practice? What is the training mechanism like?

Jane: The Fast-Slow Training, or FST, jointly optimizes those slow weights θ through reinforcement learning and fast weights ϕ through reflective prompt evolution.

Lu: They maintain a population of K textual prompts called Φ which they optimize jointly with the objective function J(θ, ϕ).

Meng: Training happens in cycles of T slow-weight updates where GEPA optimizes the fast weights Φc+one using the current policy πθc as a rollout model.

Lalam: Then for the next T steps, θ is updated on minibatches from Lc while holding Φc+one fixed.

Tom: Wait, so they are interleaving these phases? How does that affect how fast things move compared to just RL-only training?

Jane: The observation here is that the fast channel doesn't have latency because GEPA can extract task structure from a handful of rollouts and inject it through Φ immediately.

Lu: They also found that FST doesn't assume a fixed division of labor; each task draws on whichever channel pays off while still combining them when both contribute.

Meng: The in-distribution gain is decomposed into slow- and fast-weight contributions, showing that the joint cell dominates on HoVer-hard and CodeIO, while the slow weights carry most of the gain on Math.

Lalam: FST escapes the zero-reward regime by step ∼fifty which is an order of magnitude before RL begins to move signal at ∼two hundred fifty.

Tom: That means the fast channel provides enough gradient signal for FST to climb rewards quickly before the slow weights even start moving much.

Jane: The conclusion is that effective post-training should be viewed as optimization over multiple adaptation channels, where fast textual weights and slow parametric weights are trained together.

Lu: It offers a path toward post-training methods that are more data efficient, less destructive, and more amenable to continual learning.

Meng: They noted that FST reaches matched performance with one point four to three times fewer optimizer steps compared to RL-only baselines.

Lalam: The result is up to seventy percent lower KL divergence to the base policy at matched reward compared to RL-only baselines, which preserves plasticity.

Tom: So, the main promise here is data efficiency and preventing catastrophic forgetting through this split labor.

Jane: They also found that FST preserves plasticity as FST trained models adapt more effectively to subsequent tasks than parameter-only trained models.