You can deploy this model and start earning money today!

rik-codes/
dpo_humour

Name: dpo_humour
Availability: InStock

tensorboard

safetensors

Views

No views yet

Swiss-Knife

Swiss Knife: Complete Experimental Plan

Pragya Lab, BITS Pilani Goa

Paper Abstract (Swiss Knife)

Alignment today is still checkpoint-bound: most pipelines implicitly hard-code a single behavioral compromise into a trained model, making it costly to revise, specialize, or hot-fix behavior after deployment. We introduce Swiss Knife, a protocol that repurposes speculative decoding into an alignment socket: a fast draft model generates candidates, while a separately trainable auditor model enforces a chosen alignment objective during generation. Swiss Knife's key algorithmic component is the Tournament Sampling Auditor (TSA) — either a pairwise knockout bracket or a Swiss-system schedule — to select the winner under an objective-specific score. Crucially, Swiss Knife supports objective-specific auditor "blades" (safety, helpfulness, harmlessness, informativeness, style) that can be tuned, swapped, and updated independently of the backbone.

Background: What ECLIPTICA/CITA Built
Swiss Knife Architecture
Related Papers and Positioning
Experiment Plan
Key Novelty Angles
Prioritized Start Order

1. Background: What ECLIPTICA/CITA Built

ECLIPTICA is a training-time framework (Level 1) that teaches a single LLaMA-3.1-8B backbone to internalize multiple alignment regimes via natural-language instructions at inference time.

The CITA Algorithm (Contrastive Instruction-Tuned Alignment)

Training uses quadruples (I, X, Y+, Y-) where the alignment instruction I defines preference relations relative to the same prompt X.

L_CITA = L_contrastive_preference(I, X, Y+, Y−) + λ · KL(π_θ(·|I,X) || π_ref)

L_contrastive_preference: Logistic contrast on instruction-conditioned log-likelihood gaps. Features self-quenching: preference forces diminish once pairs separate.
Mandatory KL anchor: Enforces a Riemannian trust region, ensuring instruction-conditioned policies remain stable during switching.

ECLIPTICA Benchmarks and Results

Benchmark	Description	Size
ECLIPTICA	300 prompts × 10 instruction types	3,000 cases
TruthfulQA	Epistemic calibration	1,634
Conditional Safety	Policy boundary testing	1,000
Length Control	Verbosity control	1,000
LITMUS	Alignment quality index	2,800

Method	Instruction-Alignment Efficiency	TruthfulQA Adaptation
PPO	20.4%	baseline
GRPO	36.1%	baseline
DPO	56.1%	+0.001
CITA	86.7%	+0.054

Core Limitation ECLIPTICA Could NOT Solve

Backbone internalizes all regimes → they interfere with each other
Cannot hot-swap alignment objective post-deployment without retraining
New objective = full backbone retrain ($$$ cost)

Swiss Knife solves this by fully externalizing alignment into a small, swappable auditor module.

Evolution Overview

LEVEL 0: STATIC ALIGNMENT
  (DPO / RLHF / GRPO)
  - One frozen checkpoint per policy
  - No runtime policy control
  - $$$$ per new policy

        ↓  solves: "separate checkpoint per policy"

LEVEL 1: ECLIPTICA + CITA
  (Instruction-Conditioned Switching)
  - One backbone, many policies via instruction I
  - π_θ(·|I,X) trained with CITA loss + KL anchor
  - ✅ Multi-policy per checkpoint
  - ❌ Regimes interfere inside backbone
  - ❌ Cannot hot-update post-deployment

        ↓  solves: "backbone must internalize ALL regimes"

LEVEL 2: SWISS KNIFE
  (Externalized Decode-Time Alignment)
  - Draft model generates K candidates/step
  - TSA tournament selects winner under chosen objective
  - Pluggable blades: Safety | Helpfulness | Harmlessness | Style
  - ✅ Alignment fully externalized
  - ✅ Hot-swap auditors post-deployment
  - ✅ Train small auditors (~100M) independently
  - $ cost per new objective

2. Swiss Knife Architecture

Prompt X
    │
    ▼
[Draft Model — fast, frozen]  ──→  K candidates per decoding step
                                           │
                                           ▼
                            ┌─────────────────────────────┐
                            │  Tournament Sampling Auditor │
                            │           (TSA)              │
                            │                              │
                            │  Pairwise Knockout Bracket   │
                            │      OR                      │
                            │  Swiss-System Schedule       │
                            └─────────────────────────────┘
                                           │
                       ┌───────────────────┼──────────────────┐
                       ▼                   ▼                  ▼
                  [Safety Blade]  [Helpfulness Blade]  [Harmlessness Blade]
                       │                   │                  │
                       └───────────────────┴──────────────────┘
                                           │
                                           ▼
                                   Selected token/span
                                           │
                                           ▼
                                    Final Response Y

Scoring Function at Each Step

score(y) = α · log π_draft(y | x) + (1 − α) · auditor_score(y, objective)

α = 1.0 → vanilla draft decoding (no alignment)
α = 0.0 → auditor-only (ignores fluency)
α ≈ 0.4–0.6 → balanced (target operating range)

Auditor Blades

Blade	Objective	Training Signal	Auditor Size
Safety	Prevent harmful outputs	PKU-SafeRLHF, HH-RLHF	~100–300M
Helpfulness	Maximize task completion	UltraFeedback, MT-Bench	~100–300M
Harmlessness	Reduce indirect harm (bias, stereotypes)	StereoSet, WinoBias	~100–300M
Informativeness	Maximize factual accuracy	TruthfulQA pairs, FactScore	~100–300M
Style	Formality, conciseness, format	Synthetic GPT-4o pairs	~100–300M

3. Related Papers and Positioning

Direct Competitors at Decode-Time (must beat these)

Paper	Method	Why Swiss Knife Differs
SteerLM (Dong et al., 2023)	Attribute-conditioned SFT, steer via labels at inference	Still a trained backbone; no modular auditor, no tournament
Arithmetic Control (Wang et al., 2024)	Directional preference alignment with multi-objective reward vectors	Linear combination of rewards — no tournament selection, no hot-swap
Decoding-Time Multi-Obj (Shi et al., 2024)	Runtime objective weighting without retraining	Weighted sum decoding, not tournament-based; no modular blades
Inference-Time Value Guidance (Liu et al., 2024)	Value function guides tokens at inference	Continuous value signal — no discrete tournament elimination
PAD (Chen et al., 2025)	Personalized alignment at decode time	Per-user preference, not objective-type switching; no speculative decoding
CoSA (Zhang et al., 2025)	Controllable safety at inference-time	Safety-only; no general multi-blade framework
ECLIPTICA/CITA (own Level 1)	Instruction-conditioned policy trained into backbone	Cannot hot-swap; regimes interfere inside backbone

Speculative Decoding Background

Swiss Knife reframes the verifier step in speculative decoding. Instead of the verifier checking distribution match (Leviathan et al., 2023), the TSA auditor enforces a chosen alignment objective. This is the core algorithmic novelty.

Key Bibliography (119 refs in ECLIPTICA — most relevant for Swiss Knife)

Decode-time / inference-time alignment:

dong-etal-2023-steerlm — SteerLM
wang-etal-2024-arithmetic — Arithmetic Control
shi2024decodingtime — Decoding-Time Multi-Objective
liu-etal-2024-inference — Inference-Time Value Guidance
chen2025pad — PAD: Personalized Alignment at Decoding-Time
zhang2025cosa — Controllable Safety Alignment (CoSA)

Safety / red-teaming benchmarks:

mazeika2024harmbench — HarmBench: standardized red-teaming eval
ji2024pku — PKU-SafeRLHF
bai2022hh — HH-RLHF (Helpful and Harmless)
perez2022red / ganguli2022red — Red-teaming LLMs

Alignment evaluation:

borah-etal-2025-alignment — Alignment Quality Index (AQI)
lin2022truthfulqa — TruthfulQA
zheng2023judging — MT-Bench / LLM-as-judge

Preference optimization (prior work to position against):

rafailov2023direct — DPO
ouyang2022training — RLHF/PPO
meng2024simpo — SimPO
hong2024orpo — ORPO

4. Experiment Plan

Phase 0 — Setup and Baselines

Goal: Establish infrastructure and reference numbers before any novel experiments.

Model Selection

Role	Model	Rationale
Draft model	Llama-3.2-1B (or GPT-2-XL)	Fast, small, well-studied
Generator/Backbone	Llama-3.1-8B	Same as ECLIPTICA — enables direct comparison
Auditor heads	~100–300M param fine-tuned models	One per blade; cheap to train

Baseline Systems to Implement

Vanilla autoregressive — no auditor, greedy/nucleus sampling
Best-of-N sampling — generate N outputs, pick best by reward model (naive decode-time baseline)
SteerLM inference — attribute-conditioned, same prompts
ECLIPTICA/CITA — Level 1 baseline from your own prior work

Dataset Reuse from ECLIPTICA

ECLIPTICA benchmark: 300 prompts × 10 instruction types (3,000 cases)
TruthfulQA for informativeness blade evaluation
HarmBench for safety blade evaluation
PKU-SafeRLHF + HH-RLHF for auditor training

Phase 1 — TSA Core Mechanism Ablations

Goal: Prove the tournament mechanism works; find optimal hyperparameters.

Experiment 1.1 — Tournament Format Comparison

Setting	Value
Independent variable	Knockout bracket vs. Swiss-system schedule
Fixed	K=8 candidates, Safety blade, 200 prompts
Metrics	Auditor score of selected token, harmlessness score, latency (ms/token)

Hypothesis: Swiss-system is more stable under large K; knockout is faster but noisier at small K.

Expected result table:

Format	Alignment Score	Latency (ms/tok)	Stability
Knockout	TBD	Lower	Lower
Swiss-system	TBD	Higher	Higher

Experiment 1.2 — Candidate Set Size K

Setting	Value
Independent variable	K ∈ {2, 4, 8, 16, 32}
Fixed	Safety blade, knockout format, 200 prompts
Metrics	Auditor alignment score, acceptance rate (%), tokens/sec

Hypothesis: K=8–16 is the sweet spot before diminishing returns overwhelm throughput.

Experiment 1.3 — Token-Level vs. Span-Level Decoding

Granularity	Description	Trade-off
Token-level	Tournament at every single token	Higher alignment precision, more auditor calls
Span-level (n=5)	Tournament over 5-token windows	Better coherence, lower resolution

Metrics: BERTScore (coherence), auditor alignment score, auditor calls per output token.

Experiment 1.4 — Score Combination Weight α

score(y) = α · log π_draft(y|x) + (1−α) · auditor_score(y, objective)

α	Behavior
1.0	Vanilla draft — no auditor effect
0.8	Slight auditor influence
0.6	Balanced (recommended starting point)
0.4	Auditor-heavy
0.0	Auditor-only — may sacrifice fluency

Metrics: Perplexity (fluency proxy), auditor alignment score, degeneration rate (refuse-always %, boilerplate rate).

Phase 2 — Blade Training and Objective Coverage

Goal: Train and validate each auditor blade independently. Each blade is a small (~100–300M) model fine-tuned on objective-specific preference data.

Experiment 2.1 — Safety Blade

Item	Detail
Training data	PKU-SafeRLHF, HH-RLHF, HarmBench adversarial prompts
Architecture	DeBERTa-large fine-tuned as binary safe/unsafe scorer
Evaluation	HarmBench harmful compliance rate, refusal correctness
Critical test	`benign_refusal_rate` on 500 clearly benign prompts — must be <5%

Degeneration check: The auditor must NOT refuse benign queries. Measure:

benign_refusal_rate = refused_benign_count / total_benign_prompts

Experiment 2.2 — Helpfulness Blade

Item	Detail
Training data	UltraFeedback (cui2023ultrafeedback), OpenAssistant
Evaluation	MT-Bench helpfulness score, AlpacaEval win-rate vs. base model
Key test	Same prompt, swap Safety→Helpfulness blade → measure response length and information density increase

Experiment 2.3 — Harmlessness Blade

Distinct from safety: targets indirect harms (stereotypes, misinformation, implicit bias).

Item	Detail
Training data	StereoSet, WinoBias, PKU-SafeRLHF harmful preference pairs
Evaluation	CrowS-Pairs bias score, ToxiGen toxicity score
Distinction	Safety blade blocks explicit harm; Harmlessness blade reduces implicit harm

Experiment 2.4 — Informativeness Blade

Item	Detail
Training data	TruthfulQA preference pairs (truthful > hallucinated), FactScore-labeled outputs
Evaluation	TruthfulQA MC accuracy, FactScore on open-domain generation
Key insight	This blade should increase information density WITHOUT increasing harmfulness — tests decoupling

Experiment 2.5 — Style Blade (Novel Contribution)

Item	Detail
Training data	Synthetic GPT-4o generated pairs (formal vs. casual, verbose vs. concise, structured vs. prose)
Evaluation	Automated style classifiers, human preference on 100 examples
Variants	Formality sub-blade, Conciseness sub-blade, JSON/structured output sub-blade

Phase 3 — Switchability Experiments (Core Contribution)

Goal: Prove that swapping blades produces measurably different behaviors on the same prompts — directly extending the ECLIPTICA evaluation protocol to decode-time.

Experiment 3.1 — Objective Adherence Score (OA)

Protocol:

Take the 300 fixed prompts from ECLIPTICA benchmark
Run each prompt through Swiss Knife with each blade active
Score each output using all blade auditors as cross-evaluators

Metric:

OA(blade_b, prompt_x) = auditor_b(output generated with blade_b active on x)

Expected result: A 4×4 matrix where diagonal >> off-diagonal.

	Safety Auditor	Helpfulness Auditor	Harmlessness Auditor	Informativeness Auditor
Safety Blade active	HIGH	medium	high	low
Helpfulness Blade active	low	HIGH	medium	high
Harmlessness Blade active	high	medium	HIGH	medium
Informativeness Blade active	medium	high	medium	HIGH

Experiment 3.2 — Cross-Objective Separation

Metric:

Cross-Obj Separation = mean pairwise cosine distance(embedding(output_blade_i), embedding(output_blade_j))
                       averaged over all prompt pairs (i ≠ j)

Compare against:

ECLIPTICA/CITA (same prompts, different instructions)
SteerLM (same prompts, different attribute labels)
Best-of-N with same reward model

Swiss Knife should show higher separation since objectives are fully decoupled in the auditor, not merged in backbone weights.

Experiment 3.3 — Live Blade Switching (Mid-Conversation)

Setup: Multi-turn conversation where blade changes between turns.

Turn	Active Blade	Expected Behavior
Turn 1	Safety	Conservative, cautious responses
Turn 2	Helpfulness	Informative, task-complete responses
Turn 3	Safety	Returns to conservative behavior

Metrics:

Time to switch (should be ~0ms — just a module pointer swap)
Behavioral coherence: does Turn 3 behave same as Turn 1?
Cross-turn consistency score

Compare to ECLIPTICA/CITA: instruction change requires backbone re-inference; Swiss Knife requires only auditor pointer change.

Phase 4 — Robustness Experiments

Goal: Demonstrate that Swiss Knife resists degenerate solutions that plague naive reward-maximization decoding.

Experiment 4.1 — Refuse-Always Trap

Motivation: A poorly-trained safety auditor might learn to always return high scores for refusals, causing the TSA to always select refusal tokens.

Test:

Feed 500 clearly benign prompts (cooking questions, math problems, factual queries)
Measure benign_refusal_rate = # refused / 500

Comparison:

Method	benign_refusal_rate (target)
Vanilla	~1%
Greedy reward maximization (no tournament)	TBD — likely high
Swiss Knife TSA (knockout)	Target: <3%
Swiss Knife TSA (Swiss-system)	Target: <3%

Swiss Knife defense: Tournament selection eliminates single-dominating solutions; a refuse-always token wins the tournament only if refusal truly scores highest AND draft distribution supports it.

Experiment 4.2 — Boilerplate / Templating Degeneration

Motivation: Best-of-N collapses to reward-maximizing templates ("I cannot assist with...").

Metrics:

Self-BLEU = BLEU(output_i, {output_j : j ≠ i})   # lower = more diverse = less boilerplate
Distinct-1 = |unique unigrams| / |total unigrams|
Distinct-2 = |unique bigrams| / |total bigrams|

Measure over 500 diverse prompts across all blades.

Experiment 4.3 — Distribution Shift Robustness

Protocol:

Train auditors on HH-RLHF distribution
Evaluate on OOD distributions: PKU-SafeRLHF, WildGuard, AdvBench

Metric: Alignment score drop = score_in_distribution − score_OOD

Key claim: Swiss Knife's modular auditor can be cheaply fine-tuned on new distribution, unlike ECLIPTICA which requires backbone retraining.

Experiment 4.4 — Adversarial Robustness (Jailbreaks)

Attacks to run (from HarmBench):

GCG (Zou et al., 2023) — gradient-based adversarial suffix
AutoDAN — automated discrete jailbreaks
PAIR — prompt injection via iterative refinement

Metric: Attack Success Rate (ASR) — lower is better for Safety blade.

Key insight: Swiss Knife's auditor evaluates output tokens not input prompts → different attack surface than RLHF-trained models. Adversarial inputs that bypass the backbone's training are still filtered at decoding time.

Phase 5 — Systems Realism Experiments

Goal: Prove Swiss Knife is deployable under real compute constraints.

Experiment 5.1 — Acceptance Rate

Definition: Fraction of draft model tokens that survive the TSA tournament.

acceptance_rate = (# draft tokens accepted by TSA) / (# total draft tokens generated)

Target: >60% acceptance rate (comparable to vanilla speculative decoding's ~70%).

Sweep K ∈ {4, 8, 16} × format ∈ {knockout, Swiss-system}.

Experiment 5.2 — Auditor Calls Per Output Token

Theoretical complexity:

Knockout bracket: O(log K) auditor calls per token
Swiss-system (K rounds): O(K log K) auditor calls per token

K	Knockout calls/token	Swiss-system calls/token
4	2	8
8	3	24
16	4	64

Measure actual forward pass count on A100 with batch size 1 and 8.

Experiment 5.3 — Latency vs. Quality Pareto Frontier

X-axis: tokens/second throughput (on A100 80GB) Y-axis: objective-specific alignment quality score

Systems to plot:

Vanilla autoregressive (anchor point)
Best-of-N (K=8) — high quality, very slow
Swiss Knife TSA knockout (K=4)
Swiss Knife TSA knockout (K=8)
Swiss Knife TSA Swiss-system (K=8)
ECLIPTICA/CITA (instruction-conditioned, no auditor overhead)

Expected: Swiss Knife forms a better Pareto frontier than Best-of-N.

Experiment 5.4 — Auditor Size Scaling

Sweep auditor parameter count: 50M, 125M, 350M, 1.3B.

Metrics:

Alignment quality score (per blade)
Auditor inference latency (ms/token)
Quality/compute ratio

Expected: 100–300M hits the sweet spot. Above 350M gives diminishing quality gains with large latency cost.

Phase 6 — Comparison Against ECLIPTICA (Ablation Table)

Goal: The definitive paper table directly extending the evolution narrative.

Metric	DPO (Level 0)	CITA/ECLIPTICA (Level 1)	Swiss Knife (Level 2)
Instruction-alignment efficiency	56.1%	86.7%	Target: >90%
TruthfulQA adaptation	+0.001	+0.054	Target: +0.08+
ECLIPTICA benchmark OA score	—	86.7%	Target: >88%
Post-deployment hot-swap	❌	❌	✅
Regime interference	N/A	Partial (KL mitigates)	✅ None (decoupled)
Refuse-always rate (benign prompts)	~2%	~5%	Target: <3%
Self-BLEU (diversity)	high	medium	Target: low
Update cost (new objective)	Full retrain ($$$)	Backbone retrain ($$)	Small auditor ($)
Throughput vs. vanilla	1.0×	1.0×	Target: >0.7×
Cross-objective separation	low	medium	Target: high

5. Key Novelty Angles

Novelty 1 — Speculative Decoding as Alignment Socket

Nobody has reframed the verifier in speculative decoding as an alignment enforcer. Standard speculative decoding uses the verifier only to preserve the backbone's distribution. Swiss Knife repurposes this step to change the distribution toward a chosen objective. This is a fundamental reframing of what speculative decoding can do.

Novelty 2 — Tournament Sampling vs. Greedy Reward

Standard decode-time alignment methods use greedy reward maximization: argmax_y auditor(y). This collapses to degenerate solutions (refuse-always, boilerplate) because a single token can dominate. Tournament sampling via brackets:

Prevents single-token domination through elimination rounds
Combines draft likelihood signal naturally (losing bracket members have low draft score)
Formal analysis possible via multi-armed bandit / Condorcet winner theory

Novelty 3 — Objective-Specific Blade Isolation

ECLIPTICA/CITA relies on KL regularization to reduce cross-regime interference, but cannot eliminate it (all objectives share backbone weights). Swiss Knife architecturally decouples objectives into separate auditor modules. This means:

Safety updates cannot interfere with Helpfulness blade
Style blade can be added without any retraining of existing blades
Each blade can be versioned, audited, and rolled back independently

Novelty 4 — Post-Deployment Updatability

The paper makes a practical deployment claim: when a new alignment requirement emerges (e.g., stricter refusal posture for a new market), only the relevant blade needs to be retrained (~100–300M params, cheap) and hot-swapped into production. No backbone downtime, no full retraining. This is qualitatively different from anything in the current literature.

6. Where to Start (Prioritized Order)

Step 1 — Pipeline Validation (Week 1–2)

Implement TSA on top of HuggingFace's speculative decoding API.

python

# Pseudocode sketch
draft_model = LlamaForCausalLM("meta-llama/Llama-3.2-1B")
backbone = LlamaForCausalLM("meta-llama/Llama-3.1-8B")
auditor = SafetyBlade("deberta-large-safety-finetuned")

for step in range(max_steps):
    candidates = draft_model.generate(prompt, num_return=K)         # K draft tokens
    scores = auditor.score(candidates)                               # blade scores
    winner = tournament_knockout(candidates, scores, draft_scores)   # TSA selection
    output.append(winner)

Validates: The pipeline exists and produces coherent output.

Step 2 — Safety Blade Training (Week 2–3)

Fine-tune DeBERTa-large on HH-RLHF preference pairs as a binary classifier.

Easiest blade to train (clearest signal)
Most benchmarked (HarmBench, PKU-SafeRLHF)
Run Experiment 4.1 (refuse-always test) immediately after training

Step 3 — Switchability on ECLIPTICA Benchmark (Week 3–4)

Run Phase 3, Experiments 3.1 and 3.2 using ECLIPTICA's 300-prompt held-out set.

This gives you a direct comparison with CITA's published 86.7% number
Uses existing evaluation infrastructure from ECLIPTICA

Step 4 — Tournament Degeneration Test (Week 4)

Run Experiment 4.1 (refuse-always trap) formally.

This validates the central mechanical claim of TSA
Compare tournament vs. greedy reward selection directly

Step 5 — Pareto Curve (Week 5–6)

Run Phase 5, Experiment 5.3 (latency vs. quality Pareto frontier).

This is the systems realism centerpiece figure for the paper

Step 6 — Full Blade Suite + Final Ablation Table (Week 6–8)

Train remaining blades, run all benchmarks, compile Phase 6 comparison table.

Appendix: Relevant Papers Quick Reference

Citation Key	Paper Title	Relevance
dong-etal-2023-steerlm	SteerLM: Attribute Conditioned SFT	Direct baseline
wang-etal-2024-arithmetic	Arithmetic Control of LLMs	Direct baseline
shi2024decodingtime	Decoding-Time Multi-Objective Alignment	Direct baseline
liu-etal-2024-inference	Inference-Time Value Guidance	Direct baseline
chen2025pad	PAD: Personalized Alignment at Decoding-Time	Direct baseline
zhang2025cosa	Controllable Safety Alignment (CoSA)	Direct baseline
mazeika2024harmbench	HarmBench	Safety eval benchmark
ji2024pku	PKU-SafeRLHF	Safety auditor training data
bai2022hh	HH-RLHF	Safety + helpfulness training data
cui2023ultrafeedback	UltraFeedback	Helpfulness training data
lin2022truthfulqa	TruthfulQA	Informativeness eval
borah-etal-2025-alignment	Alignment Quality Index (AQI)	Alignment eval metric
rafailov2023direct	DPO	Level 0 baseline
ouyang2022training	RLHF/PPO	Level 0 baseline
dubey2024llama3	Llama 3 Herd	Backbone model
zou2023universal	GCG Adversarial Attacks	Robustness eval
perez2022red	Red Teaming LLMs	Adversarial eval
zheng2023judging	MT-Bench / LLM-as-Judge	Helpfulness eval
meng2024simpo	SimPO	Preference optimization baseline
hong2024orpo	ORPO	Preference optimization baseline

Document generated: 2026-03-24 Pragya Lab, BITS Pilani Goa Swiss Knife — Experiment Planning Document v1.0

rik-codes/dpo_humour

Swiss-Knife

Swiss Knife: Complete Experimental Plan

Pragya Lab, BITS Pilani Goa

Table of Contents

1. Background: What ECLIPTICA/CITA Built

The CITA Algorithm (Contrastive Instruction-Tuned Alignment)

ECLIPTICA Benchmarks and Results

Core Limitation ECLIPTICA Could NOT Solve

Evolution Overview

2. Swiss Knife Architecture

Scoring Function at Each Step

Auditor Blades

3. Related Papers and Positioning

Direct Competitors at Decode-Time (must beat these)

Speculative Decoding Background

Key Bibliography (119 refs in ECLIPTICA — most relevant for Swiss Knife)

4. Experiment Plan

Phase 0 — Setup and Baselines

Model Selection

Baseline Systems to Implement

Dataset Reuse from ECLIPTICA

Phase 1 — TSA Core Mechanism Ablations

Experiment 1.1 — Tournament Format Comparison

Experiment 1.2 — Candidate Set Size K

Experiment 1.3 — Token-Level vs. Span-Level Decoding

Experiment 1.4 — Score Combination Weight α

Phase 2 — Blade Training and Objective Coverage

Experiment 2.1 — Safety Blade

Experiment 2.2 — Helpfulness Blade

Experiment 2.3 — Harmlessness Blade

Experiment 2.4 — Informativeness Blade

Experiment 2.5 — Style Blade (Novel Contribution)

Phase 3 — Switchability Experiments (Core Contribution)

Experiment 3.1 — Objective Adherence Score (OA)

Experiment 3.2 — Cross-Objective Separation

Experiment 3.3 — Live Blade Switching (Mid-Conversation)

Phase 4 — Robustness Experiments

Experiment 4.1 — Refuse-Always Trap

Experiment 4.2 — Boilerplate / Templating Degeneration

Experiment 4.3 — Distribution Shift Robustness

Experiment 4.4 — Adversarial Robustness (Jailbreaks)

Phase 5 — Systems Realism Experiments

Experiment 5.1 — Acceptance Rate

Experiment 5.2 — Auditor Calls Per Output Token

Experiment 5.3 — Latency vs. Quality Pareto Frontier

Experiment 5.4 — Auditor Size Scaling

Phase 6 — Comparison Against ECLIPTICA (Ablation Table)

5. Key Novelty Angles

Novelty 1 — Speculative Decoding as Alignment Socket

Novelty 2 — Tournament Sampling vs. Greedy Reward

Novelty 3 — Objective-Specific Blade Isolation

Novelty 4 — Post-Deployment Updatability

6. Where to Start (Prioritized Order)

Step 1 — Pipeline Validation (Week 1–2)

Step 2 — Safety Blade Training (Week 2–3)

Step 3 — Switchability on ECLIPTICA Benchmark (Week 3–4)

Step 4 — Tournament Degeneration Test (Week 4)

Step 5 — Pareto Curve (Week 5–6)

Step 6 — Full Blade Suite + Final Ablation Table (Week 6–8)

Appendix: Relevant Papers Quick Reference

rik-codes/
dpo_humour