src/mindflow/data.py

"""IdeaBench-style evaluation queries (Sec 5.1 proxy).

The paper's IdeaBench is built on AI Idea Bench 2025 (Qiu et al., 2025; 3,495
papers, 8 domains) which is not openly redistributable here. We build a small,
faithful proxy: one query per domain, each with a topic x_t (problem statement,
solution withheld), related works x_r (inspirational prior papers), and an
expert-written reference idea y* distilled from a real, well-known target paper.
This is a SCALED proxy of the benchmark, not the full 3,495-paper set.
"""

QUERIES = [
    {
        "domain": "CV",
        "topic": "Building a promptable foundation model for image segmentation that generalizes zero-shot to unseen objects and tasks via flexible prompts.",
        "related": [
            "Fully convolutional and Mask R-CNN style networks segment fixed, pre-defined object categories from labelled masks.",
            "Vision transformers (ViT) provide scalable image encoders pre-trained on large image corpora.",
            "CLIP aligns images and text with contrastive pre-training, enabling open-vocabulary recognition.",
            "Interactive segmentation methods refine masks from user clicks but are trained per-dataset.",
        ],
        "reference": {
            "title": "Segment Anything: a promptable segmentation foundation model",
            "problem": "Segmentation models are trained per task/category and do not generalize to new objects or prompts; there is no foundation model for segmentation nor a web-scale mask dataset.",
            "method": "A promptable model (image encoder + prompt encoder + fast mask decoder) trained to output valid masks for point/box/text prompts; a data engine iteratively co-develops the model and a 1B-mask, 11M-image dataset via model-in-the-loop annotation.",
            "evaluation": "Zero-shot transfer to 23 segmentation datasets and downstream tasks (edge detection, proposals, instance segmentation); compare mask quality and human ratings vs interactive and supervised baselines.",
        },
    },
    {
        "domain": "NLP",
        "topic": "Aligning language models with human preferences without the instability and complexity of reinforcement learning from a separate reward model.",
        "related": [
            "RLHF fits a reward model to human comparisons and optimizes the policy with PPO, which is unstable and sample-inefficient.",
            "The optimal RLHF policy has a closed form: a reweighting of the reference policy by the exponentiated reward.",
            "Bradley-Terry models express pairwise preference probabilities via a latent scalar score.",
        ],
        "reference": {
            "title": "Direct Preference Optimization: your language model is secretly a reward model",
            "problem": "RLHF is complex and unstable, requiring reward-model training plus RL. Can we optimize the same objective directly from preference pairs?",
            "method": "Reparameterize the RLHF objective so the reward is expressed via the policy's log-ratio to a reference model, yielding a simple classification-style loss on preference pairs (no reward model, no RL, no sampling).",
            "evaluation": "Sentiment control, summarization and single-turn dialogue; compare win-rate vs PPO-RLHF and preferred-FT baselines under GPT-4/human judgment; measure stability and compute.",
        },
    },
    {
        "domain": "Multi-modal",
        "topic": "Giving a large language model general-purpose visual understanding and instruction-following using limited multimodal supervision.",
        "related": [
            "CLIP/ViT encoders yield strong transferable image features aligned to text.",
            "Instruction tuning on GPT-generated data makes LLMs follow diverse natural-language instructions.",
            "Flamingo/BLIP-2 connect frozen vision encoders to LLMs with learned adapters for captioning and VQA.",
        ],
        "reference": {
            "title": "Visual Instruction Tuning (LLaVA)",
            "problem": "There is little multimodal instruction-following data and no simple recipe to turn an LLM into a general visual assistant.",
            "method": "Use a language-only GPT to synthesize multimodal instruction-following data from image captions/boxes, then connect a CLIP visual encoder to an LLM via a projection and instruction-tune end-to-end.",
            "evaluation": "A synthetic multimodal instruction benchmark plus ScienceQA; compare against BLIP-2/visual chat baselines with GPT-4 as judge and accuracy.",
        },
    },
    {
        "domain": "Audio & Speech",
        "topic": "Endowing a large language model with perception and understanding of general audio, speech, and music from raw waveforms.",
        "related": [
            "Whisper provides robust multilingual speech recognition from weak supervision at scale.",
            "Audio encoders (BEATs) learn general acoustic representations for non-speech sounds.",
            "Q-Former style adapters bridge frozen encoders and LLMs for cross-modal reasoning.",
        ],
        "reference": {
            "title": "SALMONN: towards generic hearing for large language models",
            "problem": "Speech LLMs handle only speech; there is no single model that jointly understands speech, audio events and music and generalizes to unseen audio tasks.",
            "method": "Fuse a Whisper speech encoder and a BEATs audio encoder via a window-level Q-Former into an LLM; train with cross-modal instruction tuning plus an activation-tuning stage to unlock emergent cross-modal abilities.",
            "evaluation": "Speech recognition/translation, audio captioning, music QA and unseen tasks (e.g., audio-based storytelling); compare to task-specific and cascaded baselines.",
        },
    },
    {
        "domain": "Robotics & Control",
        "topic": "Learning visuomotor manipulation policies that model multimodal, temporally-consistent action distributions from demonstrations.",
        "related": [
            "Behavior cloning with unimodal Gaussian or MSE heads struggles with multimodal demonstrations.",
            "Diffusion models learn expressive multimodal distributions via iterative denoising.",
            "Action-chunking and receding-horizon control improve temporal consistency in imitation.",
        ],
        "reference": {
            "title": "Diffusion Policy: visuomotor policy learning via action diffusion",
            "problem": "Imitation policies collapse multimodal action distributions and produce temporally inconsistent actions, limiting dexterous manipulation.",
            "method": "Represent the policy as a conditional denoising diffusion process over action sequences, using receding-horizon control, visual conditioning, and time-series diffusion for stable, multimodal action generation.",
            "evaluation": "15 tasks across simulation and real robots; compare success rate vs LSTM-GMM/BET/IBC behavior-cloning baselines and ablate horizon and conditioning.",
        },
    },
    {
        "domain": "Science",
        "topic": "Generating novel, stable inorganic crystalline materials with targeted properties beyond substitution-based search.",
        "related": [
            "Graph neural network potentials predict formation energy and stability of crystals.",
            "Diffusion and flow models generate molecules and conformers in continuous space.",
            "Large DFT-relaxed datasets provide stability labels for inorganic compounds.",
        ],
        "reference": {
            "title": "A generative model for inorganic materials design (MatterGen)",
            "problem": "Materials discovery relies on screening known/substituted structures, rarely proposing genuinely novel stable crystals with desired properties.",
            "method": "A diffusion model over atom types, coordinates and the periodic lattice that generates stable, diverse crystals, with property-conditioned adapters fine-tuned for chemistry, symmetry, and mechanical/magnetic/electronic targets.",
            "evaluation": "Fraction of stable-unique-novel structures verified by DFT; property-conditioned generation success and distance-to-hull vs substitution and prior generative baselines.",
        },
    },
    {
        "domain": "General ML",
        "topic": "Designing a sequence model with Transformer-level quality but linear-time inference and long-context efficiency.",
        "related": [
            "Transformers have quadratic attention cost and a growing KV cache with sequence length.",
            "Structured state-space models (S4) capture long dependencies with linear recurrence but weaker language modeling.",
            "Gating and input-dependent dynamics improve selectivity in recurrent models.",
        ],
        "reference": {
            "title": "Mamba: linear-time sequence modeling with selective state spaces",
            "problem": "Attention is inefficient for long sequences while prior SSMs underperform Transformers on information-dense modalities like language.",
            "method": "Make SSM parameters input-dependent (selective) so the model propagates or forgets state per token, and implement a hardware-aware parallel scan for a simplified attention-free architecture.",
            "evaluation": "Language modeling perplexity and downstream tasks vs Transformers/Transformer++ at matched size, plus DNA and audio; measure throughput and scaling with context length.",
        },
    },
    {
        "domain": "Theory",
        "topic": "Determining how to jointly allocate model size and training tokens under a fixed compute budget for large language models.",
        "related": [
            "Kaplan et al. scaling laws suggested favoring larger models over more data.",
            "Compute-optimal training seeks the best loss for a fixed FLOP budget.",
            "Power-law fits relate loss to parameters and data separately.",
        ],
        "reference": {
            "title": "Training compute-optimal large language models (Chinchilla)",
            "problem": "Given a fixed compute budget, how should parameters and training tokens be traded off? Prior laws over-sized models relative to data.",
            "method": "Fit a parametric loss surface L(N,D) from >400 runs across model/data scales using three complementary estimation approaches, then solve for the compute-optimal (N,D) frontier.",
            "evaluation": "Predict and verify a compute-optimal model (Chinchilla) trained with the derived N,D; compare downstream accuracy vs larger under-trained models (Gopher/GPT-3) at equal compute.",
        },
    },
]


def get_queries(domains=None):
    if domains:
        return [q for q in QUERIES if q["domain"] in domains]
    return list(QUERIES)