REVIEW 38 cited by
Learning by Distilling Context
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
Language models significantly benefit from context tokens, such as prompts or scratchpads. They perform better when prompted with informative instructions, and they acquire new reasoning capabilities by generating a scratch-pad before predicting the final answers. However, they do not \textit{internalize} these performance gains, which disappear when the context tokens are gone. Our work proposes to apply context distillation so that a language model can improve itself by internalizing these gains. Concretely, given a synthetic unlabeled input for the target task, we condition the model on ``[instructions] + [task-input]'' to predict ``[scratch-pad] + [final answer]''; then we fine-tune the same model to predict its own ``[final answer]'' conditioned on the ``[task-input]'', without seeing the ``[instructions]'' or using the ``[scratch-pad]''. We show that context distillation is a general method to train language models, and it can effectively internalize 3 types of training signals. First, it can internalize abstract task instructions and explanations, so we can iteratively update the model parameters with new instructions and overwrite old ones. Second, it can internalize step-by-step reasoning for complex tasks (e.g., 8-digit addition), and such a newly acquired capability proves to be useful for other downstream tasks. Finally, it can internalize concrete training examples, and it outperforms directly learning with gradient descent by 9\% on the SPIDER Text-to-SQL dataset; furthermore, combining context distillation operations can internalize more training examples than the context window size allows.
Forward citations
Cited by 38 Pith papers
-
Can a Language Model Learn Facts Continually in Its Weights?
Facts written into LLM weights are stored but question-keyed; broad data creates usable knowledge while later writes destroy access rather than storage, leaving context more reliable than weights for composition and r...
-
Epistemic Goggles: A Pretrained Module that Induces an Epistemic Frame via Gradient Editing
Goggles is a gradient-editing module trained once per base model and frame that, when applied frozen during finetuning, causes LLMs to treat unannotated documents with a specified epistemic stance (e.g., as fiction) a...
-
Doc-to-Atom: Learning to Compile and Compose Memory Atoms
Doc-to-Atom decomposes documents into composable micro-LoRA adapters selected by a query router for efficient long-context QA.
-
Do Language Models Need Sleep? Offline Recurrence for Improved Online Inference
A sleep mechanism with N offline recurrent passes consolidates context into fast weights, improving performance on reasoning tasks where standard transformers fail.
-
From Generic Correlation to Input-Specific Credit in On-Policy Self Distillation
Self-distillation token rewards measure input-response-feedback pointwise mutual information, and CREDIT extracts the input-specific component with contrastive baselines to improve LLM reasoning performance.
-
Unmasking On-Policy Distillation: Where It Helps, Where It Hurts, and Why
Distillation signals align better with ideal updates on incorrect student rollouts than correct ones, with optimal teacher context depending on student capacity and task.
-
CoDistill-GRPO: A Co-Distillation Recipe for Efficient Group Relative Policy Optimization
CoDistill-GRPO lets small and large models mutually improve via co-distillation in GRPO, raising small-model math accuracy by over 11 points while cutting large-model training time by about 18%.
-
Near-Future Policy Optimization
NPO uses a policy's own near-future checkpoint as auxiliary trajectories to maximize effective learning signal S = Q/V, improving performance from 57.88 to 63.15 on Qwen3-VL-8B-Instruct with GRPO while accelerating co...
-
Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models
LLMs trained on simple specification gaming generalize to zero-shot reward tampering including rewriting their own reward function.
-
Learning What to Remember: Test-Time Training via Context Distillation
IP-TTCD distills the hidden-state gap between a long-window teacher and a short-window student into MLP fast weights, improving long-context language modeling and retrieval over DeltaNet, Gated DeltaNet, SWA, and IP-TTT.
-
Flux-OPD: On-Policy Distillation with Evolving Contexts
Evolving contexts improve open-ended on-policy distillation when injected as conflict-weighted corrections into a stable context-free teacher rather than used as direct targets.
-
Listen, Do Not Copy: Internalizing Audio-Grounded Scaffold Context for Robust Omni-Model Speech Understanding
Training three omnimodel speech systems with partial, audio-derived scaffold clues that are removed at test time cuts no-clue mpWER on overlapping noisy speech from 25–71% to 9–15%.
-
Sample-Efficient Learning from Agent Experience
Agents can consolidate in-context trial-and-error learning into their weights by distilling one-step teacher decisions at recorded histories, needing no additional environment interaction.
-
Can a Language Model Learn Facts Continually in Its Weights?
Weight writes create question-keyed knowledge that later sequential writes redirect, so context—not weights—is the reliable channel for facts that must survive or be composed.
-
What to Keep, What to Forget: A Rate--Distortion View of Memory Compaction in LLMs and Agents
KV-cache eviction, prompt compression, recurrent state bounding, and agent memory consolidation are unified as one rate-distortion problem with a shared lower bound, shared failure mode, and transferable mechanisms.
-
Do Modules Stay in Their Lane? Role Drift in Compound LLM Systems
RL-trained compound LLM systems can gain accuracy by having modules silently abandon their assigned roles, and a prompt-contrast regularizer can measure and limit that drift.
-
Distill to Detect: Exposing Stealth Biases in LLMs through Cartridge Distillation
D2D distills distributional shifts between a suspected model and its base into a cartridge adapter to amplify and detect stealth biases in LLMs across multiple types.
-
LiteOdyssey: A Lightweight Reasoning AI Agent for Interpretable Rare-Disease Diagnosis
A clinician-audited diagnostic policy plus public tools lets a single unmodified LLM reach high phenotype-first rare-disease Recall@1 and modestly beat baselines on real UDN patients.
-
PRISMR: Overcoming Parse Collapse in Multimodal Listwise Ranking via Parameterized Representation Internalization
PRISMR replaces in-context list processing with a hypernetwork-generated instance-specific LoRA adapter to reduce parse collapse and improve multimodal listwise ranking performance.
-
HMARS: A Hierarchical Multi-Agent Memory System for Long-Context Reasoning
HMARS introduces a hierarchical multi-agent memory system that outperforms standard retrieval and other baselines on long-document and multi-turn reasoning tasks through improved evidence coverage.
-
ThinkSwitch: Context Distillation with LoRA and Weight Interpolation for Specific-Purpose Reasoning Tasks
ThinkSwitch uses iterative self-distillation with QLoRA and spherical weight interpolation to raise both instruct and thinking checkpoint accuracy on small AIME and PubMedQA sets using only 15 human prompts per domain.
-
Tailoring Teaching to Aptitude: Direction-Adaptive Self-Distillation for LLM Reasoning
DASD improves math reasoning in LLMs by adaptively directing self-distillation based on per-token entropy to balance exploration and step accuracy, outperforming prior self-distillation and RLVR baselines on six benchmarks.
-
Context Memorization for Efficient Long Context Generation
Attention-state memory externalizes long prefixes into a lightweight lookup table of precomputed attention states, yielding higher accuracy than standard in-context learning at fixed memory budgets and lower latency t...
-
Self-Supervised On-Policy Distillation for Reasoning Language Models
SSOPD converts intra-group correct-wrong contrast into process supervision by distilling a teacher distribution from the shortest correct completion into prefixes of the longest wrong completion, improving GRPO on AIM...
-
VSPO: Vector-Steered Policy Optimization for Behavioral Control
VSPO samples rollouts at varying steering intensities to improve behavioral control in LLMs while preserving task accuracy.
-
Rethinking Dense Sequential Chains: Reasoning Language Models Can Extract Answers from Sparse, Order-Shuffling Chain-of-Thoughts
Reasoning language models extract answers from sparse, order-shuffled chain-of-thought traces with little accuracy loss.
-
TSUBASA: Improving Long-Horizon Personalization via Evolving Memory and Self-Learning with Context Distillation
TSUBASA improves long-horizon personalization in LLMs via dynamic memory evolution for writing and context-distillation self-learning for reading, outperforming Mem0 and Memory-R1 on Qwen-3 benchmarks while reducing t...
-
The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions
Training LLMs on data that enforces priority levels for instructions makes models robust to prompt injection attacks, including unseen ones, with little loss on standard tasks.
-
Large Language Models Can Self-Improve
A 540B-parameter LLM improves reasoning performance on GSM8K, DROP, OpenBookQA, and ANLI-A3 by fine-tuning on self-generated high-confidence CoT solutions from unlabeled data.
-
Agentic Reinforcement Learning with Self-Distilled Reward Shaping
ADRS converts a frozen model's privileged skill-conditioned token scores into calibrated, return-gated reward shaping for agentic reinforcement learning, reporting consistent long-horizon gains.
-
A Single Rewrite Suffices: Empirical Lessons from Production Skill Description Optimization
A single LLM rewrite of skill descriptions using false positive and negative cases matches manual optimization performance in production, with most other pipeline components adding little value.
-
Masked Distillation: Internalizing the Chain-of-Thought in Language Models
Masked distillation shows a student LLM can internalize its teacher's chain-of-thought on GSM8K but not on Countdown; emitting the last 30% of the trace as a 'suffix scaffold' recovers teacher-level accuracy with fewe...
-
Task-Aware Structured Memory for Dynamic Multi-modal In-Context Learning
TASM proposes a task-aware structured memory framework using task-vector compression, bipartite token merging, and a Core Memory plus Latent Bank hierarchy to enable efficient dynamic multi-modal in-context learning.
-
Amortizing Federated Adaptation: Hypernetwork Driven LoRA for Personalized Foundation Models
HyperLoRA amortizes federated LoRA adaptation via hypernetwork-generated initializations and product-space aggregation to fix structural bias and initialization lag.
-
Rethinking Continual Experience Internalization for Self-Evolving LLM Agents
Existing methods for turning LLM interaction experience into parametric skills collapse over multiple iterations; principle-level experience, step-wise injection, and off-policy teacher distillation yield more stable ...
-
Large Language Model Post-Training: A Unified View of Off-Policy and On-Policy Learning
LLM post-training is unified as off-policy or on-policy interventions that expand support for useful behaviors, reshape policies within reachable states, or consolidate behavior across training stages.
-
Tuning Qwen2.5-VL to Improve Its Web Interaction Skills
Two-stage fine-tuning of Qwen2.5-VL-32B improves success rates on single-click web tasks from 86% to 94%.
-
It Takes Two: Complementary Self-Distillation for Contextual Integrity in LLMs
SELFCI uses complementary self-distillation with two reverse KL divergences to align LLMs to contextual integrity while preserving utility, outperforming RL baselines like GRPO in agentic settings.
Discussion (0). Sign in to comment.