Topic deep dives¶
简体中文 · English
Compare multiple papers around one question. Start with mechanisms, then consider trade-offs and test your understanding across papers.
Each subtopic has one card. When a relevant paper arrives, update the comparison or conclusion and record what changed.
31 deep dives · 1061 topic entries
Search, ads, and recommendation · 7 cards
Where does retrieval coverage come from? User representations, positive examples, and negative difficulty
Candidate generation determines what downstream ranking can see. YouTube DNN uses one user vector for nearest-neighbor retrieval; Airbnb EBR ties supervision to the search–browse–booking journey; MIMA allocates multiple positives from a request to complementary interests; MDCNS changes negative difficulty and provides multi-view supervision. These alter representations, labels, interest specialization, and gradients, respectively. More retrieved candidates are not a common explanation.
Knowledge, collaborative vectors, and semantic codes: how does the same content enter different recommender interfaces?
KAR generates external knowledge for an existing recommender. NoteLLM learns a searchable compressed state from co-reading. Dual-purpose SID uses compact codes for both identity and content reconstruction. FineSID transfers fine-grained tokens and a complete interaction backbone across stages. Their alignment signals, compressed objects, and inherited states differ even when they all use LLMs or SIDs.
DIN, HSTU, LONGER, SlimPer, and RankMixer: what interacts with what?
Before comparing ranking backbones, trace history-to-history, candidate-to-history, and field-to-field edges. DIN activates history by candidate; HSTU updates historical states across layers; LONGER merges tokens before a short query reads long KV; SlimPer repeatedly reads the user pool but propagates only a small knowledge base; RankMixer replaces heterogeneous field interactions. These mechanisms may occupy different stages in one pipeline. A single length variable or local leaderboard cannot rank them all.
Generating an item crosses three gates: codebook, model distribution, and legal catalog
TIGER converts items to hierarchical codes and learns the next code sequence. COBRA adds a conditioned dense vector after sparse codes. PLUM continues pretraining an existing LLM in behavior and SID space. STATIC changes how legal decoding executes without changing representation; Reward Guided Decoding tilts existing support toward business preferences. Locate the changed gate before assuming an omitted item can be recovered. PinGER separates historical inputs from SID supervision and treats prefix diversity, hierarchical DPO, and catalog legality separately.
From CTR to satisfaction and uplift: define the supervision before sharing parameters
ESMM constrains click-to-conversion funnel probabilities. MMoE and PLE decide how tasks share experts. FLVM distinguishes consumption, active engagement, and positive versus negative direction. Incrementality compares potential outcomes with and without an action. More prediction heads, more shared parameters, or extra latent factors do not establish causal identification.
From slate marginal value to bid control: policy weights and constraints
CDM distills teacher choices balancing relevance and redundancy. SORT-Gen estimates prefix-conditioned marginal slate value. PRL-PUTS sweeps multi-objective policy weights. Primal-Dual Decoding updates multipliers using constraint slack during decoding. They change list context, utility modeling, policy selection, and constraint control respectively. More diversity or improvement on one metric is not an overall win. RLBP applies multi-objective governance to daily PID parameter choices, whose BU/CPC rewards and actions are not per-item slate utility.
From one cheaper forward pass to an auditable long-term research loop
ROCS removes duplicate computation within a request. Rec-Distill moves expensive teacher knowledge into training. Effective Training Time tracks training lifecycle losses. FORTE controls feature retirement risk. Self-Evolving RecSys links proposals, experiments, and online feedback over time. They optimize different timescales; lower GPU cost, feature removal, and automated proposals are not evidence of recommendation gains.
Foundation models · 6 cards
Long context has several axes: position, access, and state compression
Ask first how history is stored and then how the current token reads it. RoPE changes relative position relationships; Mamba writes history into recurrence; NSA splits access across paths; OpenAttention retains proxies for unread content; and TTR fits an input-conditioned compression. MERLIN keeps routing-key moments beside compressed segment states to change how segments are read without replacing their internal recurrent writes. A single supported-length number cannot rank these designs.
Conditional compute: choose experts, skip work, or retrieve stored knowledge?
Parameter capacity, active computation per token, and serving latency are distinct. Switch selects experts; Engram and Memory Grafting retrieve local patterns. AdaDSF, Telescopic LM, and WIDE allocate work by token and layer, full depth, and group width respectively. Identify the routing unit before discussing capacity or savings.
Training budgets: change the data, the update direction, or the comparison?
Better training within one budget can come from data mixtures, per-example editing, update quality, or lower state costs. DoReMi, Mixing Laws, DataOrchestra, Muon, and ECO act at different points. Muon Paired Budget freezes model selection and asks equal-data and equal-time questions separately. ProLand changes LoRA's product-space update geometry; GAVAGAI adds explicit word-region alignment to captioning. They are not the same intervention merely because both affect training.
Find the inference bottleneck first: weights, KV content, or cache tiers
AWQ reduces weight bit width; GQA shares KV heads; DropKV and TwinKV change retained content; KVMEM manages executable history views; Remote KV Tiers decides cross-tier loading. They can combine, but one compression ratio cannot explain quality, recoverability, and end-to-end speed. FoldBias makes compressed caches compatible with standard interfaces by folding static bias into equal-width keys, while future query drift still causes error.
Fewer sequential steps: draft verification versus learned iteration
Speculative decoding exchanges cheap proposals for expensive verification batches, preserving the target distribution with exact correction. Diffusion language models and data-space refinement instead change generation paths or training objectives. Ask who determines the final distribution, what each step does and keeps, and whether NFE or acceptance length becomes actual speed.
Evidence of capability: scoring accurately, observing less, and explaining predictions
Distinguish how correctness is defined, which example to test next, how judges are combined, and what an explanation says. XLGoBench, ProEval, FUSE, TCAB, and DetectorScope change different parts. Fewer observations and additive explanations do not imply stronger underlying capability. XLA from the Trenches adds a prior concern: compiled execution can silently depart from numerical program semantics. Successful compilation does not establish correct execution.
Structured Decisions · 3 cards
They all output candidate distributions. Why do their readout and caching boundaries differ?
Typed decisions constrain the output space, but avoiding long generated sentences does not prescribe a single architecture. LLM2Jev reads the language probability of option labels; System One reranking reads candidate letters at the first answer position; MJev-Spatial reuses native attention Q/K; and MetaEncoder encodes requests and candidates separately. Trace scoring dependencies before judging what can be cached, whether open retrieval is possible, or how continuous outputs are discretized.
Ranking invariance, calibration, and full-set recovery are distinct promises
Prism-1 constrains scoring dependencies, the JEV request audit designs paired interventions, and the multiple-correct audit checks whether a distribution recovers the correct set. They answer different questions: what the architecture guarantees, what the experiment changes, and what the evaluation measures. A valid output or one top-1 score cannot stand in for these answers.
When do extra reasoning, cross-language routing, and continual feedback change decisions?
Adaptive decisions operate on three timelines. Selective Reasoning allocates reasoning and readout within one question; a cross-language cascade routes requests across backends; Jev-DLC uses observed feedback to change later requests. Capability evaluations establish a baseline without automatically validating any control strategy. Freeze or make choices before revealing the current test target.
Multimodal models · 3 cards
What does multimodal alignment align: vectors, visual interfaces, or evidence dependence?
CLIP and EmbeddingGemma 2 produce comparable vectors. BLIP-2 and LLaVA connect vision to conditional generation, CV-OPD trains visual evidence use, and UMRM scores responses. Their supervision and interfaces differ, not a single capability ladder.
Look less, later, or again: place the visual budget at the right stage
Visual budgets are more than token counts. StackTok selects encoded evidence, PACE allocates input resolution, ViMoD retains recoverable detail, Frames-on-Demand retrieves frames, AdaLoop reprocesses existing state, and OmniKVQuant lowers cache precision. Track whether evidence survives and which costs have already been paid.
Predicting the next scene is not the same as acting or preserving a world
World models must separately show what state retains, how actions alter futures, how plans are evaluated, and how memory persists across segments. LeWAM and V-JEPA Policy connect prediction and actions; prediction–planning evaluation decomposes errors; WOVEN supervises visual transitions; PPMem maintains geometry; VBVR-Pro executes task checks. SWIM learns relative state-coordinate weights from offline action predictability to shape latent distances without giving deployment planners real future states.
LLM post-training · 6 cards
Preference is not one score: establish feedback format before choosing an alignment objective
DPO, KTO, SimPO, ORPO, and IPO change how feedback is used; CleanRM first asks whether feedback is reliable. References, length normalization, finite margins, and label-noise corrections are independent axes even if all distinguish better from worse responses.
Beyond the GRPO name: advantage scales, clipping units, and train–rollout mismatch
The GRPO family differs in how rewards become advantages, which likelihood ratio is clipped, and which distribution produced samples. Dr. GRPO, DAPO, GSPO, and TIS/IcePop modify different parts; not every ratio is the same off-policy correction.
Reward provenance determines where credit belongs
Process labels, final outcomes, self-confidence, and hindsight attribution are distinct evidence. Stepwise PRMs, MRTPO, TACO, and MART-GRPO obtain labels, localize reward, damp positive credit, and filter failed suffixes respectively. INTUITOR uses an intrinsic distributional signal, not external verification. CONCUR instead prioritizes reward dimensions by projecting gated auxiliary advantages onto a primary-reward order cone; it does not assign token credit over time.
Trajectory distillation: separate sampler, proposal, and divergence direction
On-policy describes supervised states; FKL/RKL/JSD describe how distributions compare there. GKD and MiniLLM establish those axes, DrOPD and AP-OPD adjust correction location or strength, and Semi-OPD and learning-signal diagnostics distinguish offline reuse from residual loss. MPOPD adds a learnable proposal. Its geometric-mixture target differs from the student's teacher-directed RKL.
Even stronger teachers can teach the wrong thing: evidence, compatibility, and preserving success
Teacher value is more than aggregate score. OPSD changes training visibility, MOPD and token routing organize multiple teachers, CSD aligns probability events, SPOD protects successful student actions, and VAD isolates corrections attributable to visual evidence. Verify these constraints independently before imitation.
Which compute changes: question selection, shorter generation, branching, or conditional replay?
Data and budget methods intervene at different points in training. Curricula reassess current difficulty, LESSER reduces selection-feature cost, STOPD randomly generates less and corrects estimates, BranchOPD adds important branches, Lightning caches teacher scores, and TAR retrieves suppressed correct modes. One fewer teacher call does not mean the entire run costs less.
Agents · 6 cards
Planning starts with when feedback arrives, then asks whether to branch
ReAct and ReWOO change when a plan receives observations. ToT and LATS organize candidate search differently. Jev Lookahead delegates successor simulation to code, while DRAGO checks formal-plan fidelity. They address distinct uncertainties: unknown steps, route choice, reachability, and specification errors. Do not rank them all by search count or perfect local scores.
Long-term experience needs storage, utility, and admission checks
MemGPT decides where information lives, MemPrice allocates fidelity within a limited window, and CMP studies how to identify utility. Voyager turns successful programs into skills, VaG manages admission risk, and SkillRise tests curation on later tasks. Fit, relevance, legitimacy, and future usefulness are four different questions. EMMA assigns structured memory to controller, agent, and skill, limiting each edit by failure scope and requiring actual use in the next attempt.
From making calls to delivering results: Tool interfaces, reliability, and collaboration contracts
Toolformer learns when to call, SWE-agent designs action interfaces, and RATED addresses reliability among relevant tools. AutoGen programs messaging, MetaGPT structures role deliverables, and World of SWEcraft checks collective output on hidden downstream dependencies. Any layer can fail; a valid call format is not end-to-end completion.
Identify what self-improvement changes: answers, experience, interfaces, and system versions
Self-Refine revises the current artifact, Reflexion changes what experience the next attempt can read, HarnessLens checks local workflow changes, Agent S4 evolves interface repositories, RSIFlow chooses the intervention layer, and AgentX brings candidates into a production loop with guardrails. More rounds are not a shared algorithm: persistent state and acceptance rules define what changes. CROP edits system prompts offline and penalizes output cost. EvalAlloc allocates SEARCH evaluations instead of repairing candidate code, estimates fitness against fixed references, and runs paired confirmation.
Interaction learning: How are credit and teacher information assigned after standardizing trajectories?
Agent Lightning turns execution into trainable transitions. GiGPO, AdaStep, and DepGPO assign credit using repeated states, return variance, and command dependencies, respectively. Frontier-OPD and ASTEROID change how teacher data reaches students. Recording a trajectory is only the entry point; objectives and information access define the algorithms.
Four evaluation questions: Was it completed, is it robust, why did it work, and what did it cost?
GAIA scores final answers. AgentPerturb tests single-target input changes; User Proxy Bench checks whether simulators bypass information gathering. AlphaDiana tracks failures and retries; HDA decomposes harness gains, and selective evaluation bounds automated judge error. High scores only become interpretable after these protocols are explicit.