MTA-OPSD lets each token pick its own teacher in self-distillation
On Qwen3 math tasks, the method beats prior on-policy self-distillation, with gains of up to 5.9 over the base model.
42
- Mentions · 7d2
- Mentions · 30d11
- All mentions22
- First seenSep 7, 2026
- Last mentioned3d ago
Timeline
22 articles · newest firstPretraining poison can backdoor Qwen3 CLI agents for RCE
Poisoning 0.1% of pretraining data led the 4B agent to about 92% attack success, and success stayed positive at 0.0005%.
58
Small coding agents gain little from code graphs past one hop
On SWE-bench Verified, structured tools added 17–21 points at 1.7–4B but only 4–5 points at 32B.
46
Trajectory2Gym builds decision-grounded supervision for tool-use agents
Across five Qwen3 scales it lifts the equal-weight Tau1, Tau2 and BFCL mean by 16.9 to 29.8 points.
44
LEMS and KFAC-SVD improve SVD compression of large language models
Zero-shot gains reach 4.8 points through 70B models, and KFAC-SVD lowers perplexity 15% on average.
44
TokenGate learns token pruning for faster LLM prefill
It reports 1.4–11.3× faster time-to-first-token and 1.5–3.4× lower peak memory with no accuracy loss.
46
Prompt-injection outcomes often form after the agent reads
Thinking cuts Qwen3 variance fixed at read time from 0.91 to 0.62, and single-run labels disagree up to 35%.
58
Latent Compaction with Test-Time Training
Latent Compaction preserves long-context retrieval at 32x KV-cache compression via test-time training.
48
Recovering Reliable Supervision via Critique-Guided On-Policy Distillation
CriticOPD uses teacher critiques to restore recoverability in on-policy distillation, lifting two Qwen3 students on math benchmarks.
44
τ-Emotion-Bench: Exposing the Interaction Fragility of LLM Agents through Real-World User Emotions
τ-Emotion-Bench shows emotional users expose reliability gaps in tool-calling LLM agents.
51
Latent Compaction with Test-Time Training
Latent Compaction keeps 96.7% RULER accuracy at 32x KV-cache compression via test-time training.
56
Probe-Space Preconditioning for Fast and Stable Zero-Order Training
1.5-SPSA preconditions zero-order probes so OPT-13B beats MeZO on SST-2 in 70 steps instead of 100,000.
58
Inducing Process Supervision from Outcome-Only Reinforcement Learning
TIPS trains generative process reward models with outcome-only RL, reaching 85.2 F1 on ProcessBench.
58
ReSPO: Reshaped Sequence Policy Optimization for Gradient Starvation in Off-Policy Learning
ReSPO replaces clipping in off-policy RL to fix gradient starvation on reused rollouts for Qwen3.
44
DuoOPD: Learning from Joint Teacher-Student Outcomes for Multi-Task On-Policy Distillation
DuoOPD uses joint teacher-student outcomes to improve multi-task on-policy distillation over standard OPD.
49
Pruned CTC for Memory-Efficient Large-Vocabulary ASR Training
Pruned CTC restricts alignment computation to target tokens and blank, cutting ASR training memory 5.1x for 180K-vocabulary LLMs with minimal accuracy loss.
24
Receiver-Conditioned Latent Communication gives 94% CacheBack
CacheBack filters multi-agent KV caches by receiver need, cutting 75% of state and raising FanOutQA accuracy 14.7 points.
48
When EOS Tokens Disagree: Understanding Length Inflation in On-Policy Distillation
Researchers show EOS termination-token mismatch between base students and post-trained teachers drives length inflation in on-policy distillation across Qwen3, Llama, and Gemma.
28
A Zeroth-Order Paradigm for LLM Preference Alignment
Researchers propose ComPO, a zeroth-order comparison-based preference alignment method with convergence guarantees that mitigates likelihood displacement in LLMs.
32
A Zeroth-Order Paradigm for LLM Preference Alignment
ComPO is a zeroth-order preference alignment method using comparison oracles to mitigate likelihood displacement across Mistral, Llama, Gemma, and Qwen3 models.
38
MoME: Mixture-of-Memory Embeddings for Context-Aware Sparse Lookup
MoME introduces context-aware memory embeddings with gated multi-slot lookup, improving LLM efficiency across Llama-3 and Qwen3 backbones.
30
Environments as Scaffold: Enriching Feedback to Bootstrap Self-Evolving Agents in Long-Horizon Tasks
Researchers propose Feedback-Enriched Environments (FEEs) that reduce reward sparsity and improve RL training of Qwen3-based agents on SciWorld and BFCL.
25
Appears with
- Llamaai model · 4
- Mistralai model · 3
- Gemma-2ai model · 2
- MiniCPM5ai model · 2
- THUNLPorganization · 2
- Gemma-3ai model · 2
- nanochatproduct · 1
- OPT-30Bai model · 1
- Qwen3.5ai model · 1
- AgentDojoproduct · 1
- τ-Emotion-Benchproduct · 1
- Alibabavendor · 1
- ProcessBenchproduct · 1
- Qwen3product · 1
- Llama-3ai model · 1
- τ-Benchproduct · 1
Entities are extracted by the model from each article. First seen , last mentioned . Watching keeps it in this browser only (no account); the watchlist page and dashboard alerts use it.