Get Started
Topic · 77 recaps
LLM Pretraining & Post-training
How large language models are built: data curation and mixtures, pretraining objectives, scaling laws, instruction tuning, preference learning, and the full post-training stack.
Play all
...
Posts
Questions
Home
Topics
Search
Library
Questions researchers are working on
Follow a question through Rcap’s explanations and the latest papers addressing it.
Search
Can automated alignment research mitigate multiple measurable safety failures without sacrificing general model capability?
Alignment failures such as deception, sycophancy, and jailbreaks can be measured, but reducing several simultaneously may interfere with a model’s broader capabilities. It is also unclear whether automated researchers can develop effective interventions without extensive human guidance.
Can black-box attackers identify and reconstruct prompts supposedly removed from language models without knowing them in advance?
Machine unlearning may suppress responses to removed data without eliminating signals that reveal what was removed. The difficulty is determining whether an attacker can use those signals to discover and reconstruct forgotten prompts that are initially unknown.
Can large language models reliably perform Arabic morphosyntactic tagging and dependency parsing despite morphological and orthographic ambiguity?
Arabic’s rich morphology and orthographic ambiguity make morphological and syntactic interpretation closely interdependent. Performance can also vary with how text is represented and whether relevant annotated examples are available as demonstrations.
Can post-training ternarization make language models smaller without unacceptable capability loss or slower inference?
Ultra-low-bit weights can shrink model storage, but nominal bit counts may not reflect the stored representation, uneven task degradation, or actual inference speed. Compression may therefore improve footprint without improving end-to-end deployment performance.
Can user feedback reliably guide LLM revisions if LLM judges overlook the resulting improvements?
User interactions may reveal issues that an LLM cannot detect on its own, but the feedback can be noisy and improvements may be difficult to measure. Evaluation becomes especially problematic when judges prefer a baseline response even after feedback has corrected the targeted issue.
Does Cantonese-specific adaptation improve language-model prediction of naturalistic Cantonese reading consistently across information-theoretic measures?
Language models estimate information-theoretic expectations that may help explain human reading, but it is unclear whether Cantonese-specific adaptation improves this alignment. Different measures can rank models differently, making conclusions sensitive to how predictive fit is quantified.
Does reusing Transformer layers improve language-model quality when parameter, compute, and KV-cache budgets are matched?
Layer looping increases effective computation by revisiting shared parameters, but comparisons can mistake extra computation or memory for an architectural improvement. The central difficulty is isolating the effect of reuse while holding major training and inference budgets constant.
How can a single search agent improve multi-hop web research without sub-agents or test-time verification?
Multi-hop web research requires an agent to connect evidence across several pages while retaining the information needed for later steps. Long search trajectories can overwhelm the available context, making both training and reliable performance difficult without additional agents or verification passes.
How can alignment systems infer the multiple criteria behind human pairwise preferences?
Pairwise preference labels record which option people choose, but not the interacting considerations behind that choice. This makes it difficult to build preference models that are both faithful to judgments and interpretable.
How can autoregressive language models add recurrent cross-token memory without breaking KV caching or one-forward-per-token decoding?
Standard transformers rely on attention and cached keys and values during generation, but they do not naturally carry a separate recurrent state across tokens. Adding one can alter the decoding interface or require additional forward passes.
How can autoregressive logit distillation preserve local next-token preferences when teacher and student rankings disagree?
Matching every vocabulary logit can underrepresent distinctions among likely alternatives. Selecting candidates from only the teacher or student can also omit useful tokens or trust unreliable early student rankings.
How can banking agents safely handle sensitive account requests while staying grounded and using tools correctly?
Banking agents must rely on trusted bank-specific information while distinguishing safe assistance from risky or out-of-scope requests. Account-related interactions also require correct tool use and cautious handling of sensitive customer situations.
How can batched LLM serving handle request-specific read-write adaptation state without corrupting streams or losing throughput?
Standard batched LLM serving assumes shared static weights, but test-time training changes request-owned state during generation. Mixing incompatible reads and writes can transfer state across streams, while serial execution limits throughput.
How can code-repair models fix bugs without unnecessarily rewriting surrounding code?
A repair can pass correctness tests yet alter unrelated code, making review harder and weakening confidence that the original implementation was preserved.
How can coding and terminal agents be post-trained without distorting production-faithful token flows and control operations?
Simplified training environments and reconstructed logs can change the prompts and token flows encountered in deployment. They can also mix control operations with ordinary model activity, causing training and evaluation to reflect the wrong execution process.
How can compact text embedding models improve retrieval and generalization through better training and data quality?
Compact embedding models can underperform on retrieval and generalization when development emphasizes data scaling or synthesis without sufficiently addressing training methods and data quality. The problem is to improve their representations without relying on substantially larger models.
How can complete transformer layers be removed from LLMs without sacrificing quality or standard serving compatibility?
Removing whole transformer layers can lower decoding cost, but quality losses vary across models and can become severe at aggressive compression levels. The challenge is to reduce model depth without requiring an altered serving architecture.
How can computer-use agents efficiently coordinate GUI and CLI actions over shared application state?
Real computer work often combines visual interface inspection with command-line operations, yet agents commonly rely too heavily on one modality. This produces inefficient GUI trajectories or brittle scripts, while environments exposing both interfaces across real applications remain scarce.
How can deep-network momentum adapt forgetting to unevenly sampled input directions?
A single exponential decay treats frequently and rarely activated directions alike. As a result, stale gradient information may persist in some directions while other directions are forgotten too quickly.
How can dense retrievers specialize for biomedical search without sacrificing general-domain retrieval?
Training retrievers across biomedical and general domains can be costly, and mixed-domain training may blur capabilities that are useful in specialized settings. The central difficulty is improving biomedical retrieval while preserving broad-domain performance.
Previous
1 / 7
Next