AI Research Weekly – autonomous driving & more – May 10, 2026

Summary

This research discovers that large language models internally organize social roles along a single dominant dimension: granularity, ranging from micro-level individual perspectives to macro-level institutional reasoning. The authors construct a 'Granularity Axis' using contrast-based methods on 75 ordered social roles, finding it aligns with the primary component of role representation space and accounts for over half the variance. Crucially, they demonstrate this axis is causally manipulable through activation steering, allowing real-time control over whether an LLM responds from an individual, community, organizational, or institutional perspective. The findings reveal that role-conditioned behavior operates along a continuous social-scale manifold rather than discrete personas.

Key findings

• Social role granularity is the dominant geometric axis in LLM role representation space, accounting for 52.6% of variance in Qwen3-8B
• A contrast-based Granularity Axis constructed from micro/macro endpoints successfully predicts intermediate granularity levels with monotonic ordering
• The axis transfers across model families (Qwen3-8B and Llama-3.1-8B) and remains stable across layers and prompt variations
• Activation steering along this axis causally shifts output granularity, enabling real-time control of social perspective scale
• Models differ in steering responsiveness based on their default operating regime, with some showing ceiling effects

How to implement

• Build multi-agent debate systems that actively monitor and prevent perspective collapse by measuring agent positions on the granularity axis during conversations
• Develop customer service chatbots that dynamically adjust social perspective based on query type - individual support mode for personal issues, institutional mode for policy questions
• Create policy simulation platforms that apply granularity steering to ensure different stakeholder agents (citizens, organizations, governments) maintain appropriate perspective scales throughout scenarios

Summary

The AI Co-Mathematician introduces a stateful workspace where multiple AI agents collaborate asynchronously on mathematical research tasks. Unlike traditional chat interfaces, it maintains persistent project state, manages uncertainty through review cycles, and produces native LaTeX documents with margin annotations. The system delegates work across specialized agents (literature review, computation, theorem proving) while allowing human steering. Early users solved open problems including a Kourovka Notebook question. The system scored 48% on FrontierMath Tier 4, surpassing previous AI systems through orchestration rather than raw model capability improvements.

Key findings

• Stateful multi-agent orchestration significantly outperforms single-model approaches, achieving 48% accuracy on FrontierMath Tier 4 versus 19% for base Gemini 3.1 Pro
• Interactive human steering is crucial - early users successfully resolved open mathematical problems through iterative collaboration with the AI system
• Native mathematical artifacts (LaTeX with margin annotations) enable better uncertainty communication and workflow integration than transient chat outputs
• Asynchronous parallel workstreams with hard programmatic constraints prevent common AI failure modes like hallucinated proofs and premature success claims

How to implement

• Build a stateful research assistant for legal teams that maintains case files, delegates document review to specialized agents, and produces annotated briefs with uncertainty markers for complex litigation workflows
• Create an engineering design workspace where AI agents handle parallel workstreams for requirements analysis, literature review, and prototype validation while maintaining persistent project state and review cycles
• Develop a scientific research platform that orchestrates AI agents for hypothesis generation, literature synthesis, and experimental design while tracking failed approaches and enabling human steering of long-running investigations

Summary

This paper introduces LOPE (Lorem Perturbation for Exploration), which prepends randomly generated Latin placeholder text to prompts during reinforcement learning training. When all generated responses fail for a difficult question (zero-advantage problem), LOPE adds Lorem Ipsum sequences to shift the model's output distribution and discover new reasoning pathways. Experiments show consistent improvements across 1.7B-7B parameter models on math reasoning tasks. The key insight is that task-irrelevant, low-perplexity perturbations can break models out of local reasoning patterns without corrupting comprehension, outperforming both standard resampling and high-temperature generation approaches.

Key findings

• Lorem Ipsum perturbations unlock orthogonal reasoning pathways that standard logit-space exploration (temperature sampling) cannot reach
• Effective perturbations require pseudo-Latin vocabulary and low perplexity to avoid interfering with English reasoning while maintaining response quality
• LOPE maintains higher question-level success rates throughout training compared to naive resampling, improving data utilization efficiency
• Training signal shaping techniques are necessary to handle off-policy optimization when mixing perturbed and original responses

How to implement

• Integrate LOPE into existing RLHF training pipelines for code generation models to improve success rates on hard programming problems by prepending Latin sequences when all initial solutions fail
• Apply prompt perturbation during fine-tuning of customer service chatbots to discover diverse response strategies for difficult queries, using controlled noise to break out of repetitive answer patterns
• Implement LOPE in mathematical reasoning systems like homework helpers or tutoring applications to generate alternative solution approaches when standard methods fail, improving problem-solving coverage

Summary

TIDE introduces a parallel memory system that maintains token identity information throughout all transformer layers, addressing two critical problems: rare tokens receiving insufficient training updates due to frequency imbalances, and semantically distinct tokens becoming indistinguishable in similar contexts. The architecture adds K independent memory blocks that map token indices to semantic vectors, injected into each layer via learned routers. This provides rare tokens with amplified gradient signals and prevents contextual collapse by maintaining discrete token-specific information independent of contextual hidden states. Experiments show consistent improvements across model scales and downstream tasks.

Key findings

• Rare tokens receive orders of magnitude fewer gradient updates than common tokens, leading to systematically undertrained embeddings
• Contextual collapse causes semantically distinct tokens to become indistinguishable in similar syntactic environments due to FFN Lipschitz constraints
• TIDE's K-pathway gradient amplification provides rare tokens with K-fold increase in training signal compared to standard transformers
• Memory blocks learn to specialize for different frequency regimes, with routers adaptively weighting memory contributions based on token rarity

How to implement

• Implement TIDE architecture in domain-specific language models handling technical terminology or rare entities (medical, legal, scientific) to improve rare term understanding and generation accuracy
• Deploy TIDE in multilingual models where low-resource languages suffer from rare token problems, using memory blocks to maintain language-specific semantic representations across all layers
• Integrate TIDE into code generation models to better handle rare API names, variable identifiers, and domain-specific tokens that appear infrequently in training data

Summary

UniPool fundamentally redesigns Mixture-of-Experts (MoE) by replacing the standard approach of giving each layer its own private expert set with a single global pool of experts shared across all layers. This addresses redundancy in deep MoE layers where experts learn similar transformations. The approach uses pool-level load balancing and NormRouter for stable training. Across five model scales (182M-978M parameters), UniPool consistently outperforms vanilla MoE while using only 41.6-66.7% of the expert parameters, enabling sublinear expert scaling with model depth.

Key findings

• Deep MoE layers exhibit substantial expert redundancy - randomizing routing in single deep layers drops accuracy by only 1.0-1.6 points
• Global expert sharing consistently improves validation loss across all tested scales (182M-978M parameters)
• Pool size becomes an explicit scaling hyperparameter - expert parameters can grow sublinearly with depth while maintaining performance
• Shared pool requires co-designed components: pool-level auxiliary loss and NormRouter for stable training

How to implement

• Deploy larger language models with reduced memory footprint by using UniPool's sublinear expert scaling to maintain quality with fewer expert parameters
• Optimize existing MoE model architectures by replacing per-layer expert sets with shared pools, potentially reducing training and inference costs by 35-60%
• Build deeper transformer models more efficiently by allocating expert capacity globally rather than linearly with depth, enabling better depth-width tradeoffs

Summary

Cola DLM proposes a hierarchical approach to language modeling that first maps text to continuous latent variables via a VAE, then models semantic structure using a block-causal diffusion transformer, and finally generates text through conditional decoding. This separates global semantic organization from local text realization, enabling non-autoregressive generation and natural multimodal extension. The key insight is using diffusion for latent prior transport rather than token-level recovery. Experiments show competitive scaling behavior compared to autoregressive baselines, though with notable likelihood-generation quality misalignment.

Key findings

• Diffusion-based latent prior transport outperforms token-level observation recovery for text generation
• Generation quality and likelihood estimation can be structurally misaligned in continuous latent language models
• Joint training of VAE and diffusion transformer from stable initialization achieves better scaling than fixed or scratch training
• The approach naturally extends to unified text-image modeling through shared continuous latent space

How to implement

• Build a document editing system that generates coherent long-form content by modeling global document structure in latent space before generating local text segments
• Develop a multimodal chatbot that processes both text queries and images by mapping both modalities to a shared continuous latent space for unified reasoning
• Create a content generation pipeline for marketing materials that first plans semantic structure (tone, key points) in latent space, then realizes text conditioned on visual brand elements

Summary

This paper introduces SCALELOGIC, a synthetic logical reasoning framework that enables precise control over reasoning depth and logical expressiveness for studying RL training efficiency. The key discovery is that training compute follows a power law T ∝ D^γ with reasoning depth, where the scaling exponent γ increases monotonically from 1.04 to 2.60 as logical expressiveness grows from simple implication-only to full first-order reasoning with quantification. More expressive training settings yield larger downstream performance gains (up to +10.66 points) on real mathematical reasoning benchmarks, demonstrating that what models train on, not just training volume, shapes transfer performance. Curriculum-based training substantially improves scaling efficiency across all settings.

Key findings

• RL training cost follows power law T ∝ D^γ with reasoning depth, where scaling exponent γ increases monotonically with logical expressiveness (1.04 to 2.60)
• More expressive training logic yields larger downstream performance gains on mathematical reasoning benchmarks (up to +10.66 percentage points)
• Curriculum training substantially improves scaling efficiency, reducing power-law exponent from 2.60 to 2.30 in most expressive setting
• Power-law scaling relationship holds across multiple RL algorithms (DAPO, GRPO, GSPO), indicating broad applicability

How to implement

• Build RL training curricula for mathematical reasoning models by starting with simple implication-only problems and progressively introducing conjunction, negation, and quantification to maximize compute efficiency
• Design synthetic data generation pipelines for code reasoning assistants using SCALELOGIC's controlled complexity framework to systematically scale training difficulty while maintaining verifiable rewards
• Optimize resource allocation for reasoning model training by using the discovered power-law relationships to predict training compute requirements based on target reasoning depth and logical complexity

Summary

A²TGPO improves reinforcement learning for multi-turn AI agents by solving three key problems in credit assignment. Instead of pooling rewards across all turns, it groups rewards by turn position, recognizing that agents at the same interaction depth face similar contexts. It rescales accumulated rewards to prevent later turns from dominating gradients, and adapts clipping ranges based on how informative each turn is. Testing on question-answering tasks with tool use shows consistent 1.7+ point improvements over existing methods across different model sizes.

Key findings

• Turn-group normalization by (prompt, turn-index) eliminates cross-position incomparability in multi-turn agent training
• Variance-rescaled discounted accumulation keeps advantage magnitudes comparable across different trajectory depths
• Adaptive turn-level clipping based on information gain allows more aggressive updates for informative turns
• Method achieves +1.75 average improvement on multi-hop tasks and +1.69 on single-hop tasks across three model backbones

How to implement

• Train customer service chatbots that use multiple tools (CRM lookup, knowledge base search, ticket creation) by properly crediting which tool calls actually help resolve issues versus just extending conversations
• Improve code generation agents that iteratively run tests, read documentation, and refactor code by better rewarding the specific debugging steps that lead to working solutions
• Optimize web navigation agents for e-commerce or form-filling tasks by distinguishing between productive clicks that advance toward goals versus exploratory actions that gather context

Summary

This paper introduces Continuous-Time Distribution Matching (CDM), which improves few-step diffusion model generation by replacing fixed discrete timestep training with dynamic continuous scheduling. Instead of training only at predetermined timesteps that match inference, CDM uses randomly sampled continuous timesteps and introduces an off-trajectory supervision mechanism. This mechanism uses velocity-driven extrapolation to create intermediate points between sampling steps and enforces distribution matching on these off-trajectory latents. The approach achieves state-of-the-art 4-step generation quality without requiring GANs or reward models, demonstrating that the common assumption of strict training-inference alignment is unnecessarily restrictive.

Key findings

• Training-inference timestep alignment is not necessary and actually restricts performance - dynamic continuous scheduling outperforms fixed discrete schedules
• Distribution matching loss captures the teacher's CFG-free distribution rather than merely acting as a training stabilizer
• Velocity-driven off-trajectory supervision effectively corrects numerical integration errors that occur during few-step inference
• CDM achieves state-of-the-art 4-step generation quality without requiring complex auxiliary objectives like GANs or reward models

How to implement

• Integrate CDM into existing text-to-image APIs to reduce inference latency from 50-100 steps to 4 steps while maintaining quality, enabling real-time image generation for interactive applications
• Apply continuous-time scheduling to custom diffusion model training pipelines for domain-specific applications like medical imaging or product visualization where speed is critical
• Implement velocity-driven extrapolation in existing diffusion distillation frameworks to improve quality of few-step video generation models for content creation platforms

Summary

SwiftI2V tackles the computational challenge of generating high-resolution (2K) videos from images by splitting the process into two stages: low-resolution motion generation followed by high-resolution detail synthesis. The key innovation is Conditional Segment-wise Generation (CSG), which processes videos in small temporal segments with bidirectional attention between segments, avoiding memory explosion while maintaining quality. This achieves 202× speedup over end-to-end approaches while matching quality, enabling 2K video generation on consumer GPUs like RTX 4090.

Key findings

• Two-stage decoupling of motion modeling and detail synthesis enables 202× GPU-time reduction compared to end-to-end 2K video generation
• Conditional Segment-wise Generation with bidirectional attention maintains quality while processing videos in bounded memory segments
• Stage transition training with synthesized artifacts bridges the gap between separately trained cascade stages
• Method achieves competitive VBench-I2V scores while reducing peak memory to 33.5GB, enabling consumer GPU deployment

How to implement

• Integrate SwiftI2V into content creation platforms to offer real-time 2K video generation from user-uploaded images without requiring expensive GPU clusters
• Deploy segmented generation approach in mobile video editing apps, processing long-form content in streaming fashion to avoid memory constraints
• Adapt bidirectional segment attention mechanism for other high-resolution generative tasks like panoramic image synthesis or long document generation