Math Discovery, Long-Context Memory, and the Limits of Multimodal Reasoning

AI-driven math breakthroughs, 4M-token reasoning, memory-first agents, why video models still fail at causality, and faster diffusion decoding.

Before we jump in, meet today’s sponsor - Sequa!
Your documentation should maintain itself. AI-native docs powered directly by your codebase.

Sequa transforms your entire codebase into a living knowledge graph that powers self-maintaining documentation and real AI understanding.

Stop rewriting Confluence pages and chasing tribal knowledge. With AI-native docs, your team - and your AI agents - always have accurate context.
Less friction, fewer errors, faster shipping.

HackerPulse readers get exclusively 50% off Sequa for 2 months. Redeem in check out by Dec 31 with code: HACKERPULSE50

---

This week’s research pushes AI deeper into mathematical discovery, long-context reasoning, and memory-centric agent design. We also get a reality check on multimodal reasoning gaps and a new decoding paradigm that blends diffusion with autoregressive efficiency.

Here’s what’s new:

🧮 Mathematical Exploration and Discovery at Scale — AlphaEvolve, an evolutionary coding agent, autonomously discovers and improves mathematical constructions across 67 problems. It generalizes solutions into universal formulas and integrates with proof assistants, marking a new era of human–AI collaboration in math research.

🎥 MMGR (Multi-Modal Generative Reasoning) — A comprehensive benchmark exposing a major gap in video and image models: while visuals look convincing, causal and abstract reasoning collapses (<10% accuracy). Highlights why today’s world simulators still fail at true reasoning.

🧠 QwenLong-L1.5 — A post-training recipe that unlocks long-context reasoning up to 4M tokens. Combines synthetic multi-hop data, stabilized RL, and memory-augmented architectures to match frontier models on long-context benchmarks while boosting general reasoning.

🗃️ Memory in the Age of AI Agents — A unifying survey that organizes agent memory into forms, functions, and dynamics. Positions memory not as an add-on, but as a core design primitive for future agentic systems.

⚙️ ReFusion — A diffusion-based language model that decodes in parallel at the slot level using a plan-and-infill process. Achieves 2.33× speedups, reuses KV cache efficiently, and closes much of the gap with standard autoregressive models.

---

Mathematical exploration and discovery at scale (🔗 Read the Paper)

AlphaEvolve, an evolutionary coding agent combining LLMs with automated evaluation, autonomously discovers novel mathematical constructions across 67 problems in analysis, combinatorics, geometry, and number theory, rediscovering known solutions and finding improvements while demonstrating the potential for AI-guided mathematical discovery at scale. The system can generalize finite solutions to universal formulas and integrate with proof assistants to provide automated verification and insights, establishing a new paradigm for human-AI mathematical collaboration.

---

MMGR: Multi-Modal Generative Reasoning (🔗 Read the Paper)

MMGR introduces a comprehensive evaluation framework that assesses video and image foundation models on five reasoning abilities (physical, logical, 3D/2D spatial, temporal) across abstract reasoning, embodied navigation, and physical commonsense tasks, revealing that current state-of-the-art models achieve strong perceptual quality but fail dramatically at causal reasoning and global consistency (e.g., <10% accuracy on abstract reasoning tasks). This benchmark exposes a critical gap between visual plausibility and reasoning correctness in generative models, establishing metrics and diagnostics needed to develop more reliable world simulators.

---

QwenLong-L1.5: Post-Training Recipe for Long-Context Reasoning and Memory Management (🔗 Read the Paper)

QwenLong-L1.5 achieves state-of-the-art long-context reasoning through three innovations: a data synthesis pipeline generating multi-hop reasoning tasks, stabilized reinforcement learning with adaptive entropy control for long-context training, and a memory-augmented architecture enabling reasoning over 4M-token sequences. The model matches GPT-5/Gemini-2.5-Pro performance on long-context benchmarks while improving general reasoning capabilities.

---

Memory in the Age of AI Agents (🔗 Read the Paper)

This survey provides a comprehensive taxonomy of memory systems in AI agents, organizing the fragmented literature through three unified lenses - forms (token-level, parametric, latent), functions (factual, experiential, working), and dynamics (formation, evolution, retrieval)- while identifying key research frontiers and positioning memory as a fundamental design primitive for future agentic systems.

---

ReFusion: A Diffusion Large Language Model with Parallel Autoregressive Decoding (🔗 Read the Paper)

ReFusion introduces a masked diffusion model that achieves both superior performance and efficiency by decoding at the slot level (fixed-length sub-sequences) rather than individual tokens, using an iterative “plan-and-infill” process that enables full KV cache reuse and reduces learning complexity while maintaining 2.33× speedup over autoregressive models. The approach outperforms prior diffusion models by 34% while closing the performance gap to standard autoregressive models.

---

And before you leave, watch this demo from Sequa and try it our for 50% off.

HackerPulse readers get exclusively 50% off Sequa for 2 months. Redeem in check out by Dec 31 with code: HACKERPULSE50